Home

Project Report - Duncans McPherson`s Home Page

1. 2 0 02 ee eee eee The same IO space display in 3d from 3 different viewpoints Animation of changing IO space during supervised learning The source files comprising JSAND llle An example subgoals applet display The main JSAND application window The JSAND application fle meni s sop s ecs aie spoe abe m c c The Wintermute display mode nooo lll An Island mode display of two hyperbands and a boundary line An jsand mode display of two hyperbands a The subgoal display applet in a Solaris desktop iv o Noaoa AN 10 11 12 13 14 17 24 37 37 39 40 43 44 46 54 55 56 57 57 58 Chapter 1 Problem Definition 1 1 Problem Definition There are two main objectives of this project One is the generation of a real time three dimensional representation of the Input Output I O mapping of an Artificial Neural Net work An I O mapping represents the complete set of output patterns of a neural network for all possible combinations of input pattern according to the current weight state of that network This project will allow a network s current I O mapping to be rendered in real time while training of that network is underway The second objective is to allow the path of the output behaviour of the network to be rendered in output space This representation is then to be extended to cater
2. description display network loadNetwork is while description null 1 display network generateNetwork description if display network loadedCorrectly display bands display network calculateHyperbands cb 12 2 The JSAND applet JsandApplet java Software Listings display regions display network calculateRegions display boundaries display network calculateBoundaries display regions if display region activation display regions display network calculateWintermute display translated null display planes display network calculatePlanes units setText new String display network Hidden length display repaint try Thread sleep 500 catch InterruptedException e description display network loadNetwork is else description null display network null fin close catch Exception e System out println File error during load e 12 2 LIII The JSAND applet JsandApplet java Eus db dee E K k ak epo SEO SORGE IE HOO oe JSand Hyperband display system applet build K KK import import import import import import import import By Duncan McPherson Sth of March 2000 E hae k k ak K SE K K ak ae K ak ak k ERE SIEGE K ak ae ak K K k k k java awt event java awt java io java util java awt image java applet java io java net D 12 2 T
3. Button ok button The button to dispose of the dialog box Panel panel The panel containing the text public AboutDialog Frame parent_frame String title super parent_frame title false setSize 320 225 addWindowListener this panel new Panel panel add new Label Artificial Neural Network Label CENTER 68 12 1 The Jsand Application Jsand java Software Listings panel add new Label Input Output Space Display Label CENTER panel add new Label By Duncan McPherson Label CENTER panel add new Label Senior Honours Project April 2000 Label CENTER ok button new Button Close About Box ok button addA ctionListener this add South ok button add Center panel setLocation 200 150 show public void actionPerformed ActionEvent event if event getSource ok_button dispose public void windowClosed WindowEvent event dispose public void windowClosing WindowEvent event dispose public void windowDeiconified WindowEvent event public void windowlIconified WindowEvent event public void windowActivated WindowEvent event public void windowDeactivated WindowEvent event public void windowOpened WindowEvent event class EditDialog extends Dialog implements ActionListener WindowListener Button ok close The button to dispose
4. Weight inputWeights elementAt 1 weight result gradient wlx wly result top k wly biasWeight wly result bot k wly biasWeight wly return result j j Represents the weight value that connects two neurons class Weight 1 public Neuron from public Neuron to public double weight public Weight Neuron efferent Neuron afferent double value from efferent to afferent weight value from output Weights addElement this public Weight Neuron efferent Neuron afferent Random generator 95 12 4 The hyperband data storage module hBand java Software Listings Initialise all values pertaining to the weight object from efferent to afferent weight 1 generator nextDouble 2 Theinput neuron to this weight line has the weight added to it from output Weights addElement this 12 4 The hyperband data storage module hBand java CK FR FOO I I IR k k ak ak K k k k K K k Hyperband data structure module By Duncan McPherson x This module is used to represent a hyperband k k k k k ak ak k ak ak K k ak K ak ak k k ak K k ak K k ak 2K k ak ak k ak K K ak K K k k K k class hBand public double gradient 1 0 public double width 20 0 public double top 0 0 public double bot 0 0 12 5 The input output display module I0SpaceCanvas java JS Eoo d ICH k k k OE k k k k
5. 37 7 1 2 The Island display mode 002000000 38 7 1 8 The Jsand display mode 2 2 00004 40 7 1 4 The load and save functions of the JSAND Application 41 7 1 5 The JSAND applet in client mode 42 GLO Implementation of JSAND 222949 Rr ee ee ee ee eS 44 12 The subsoal chain display si oe coia eae wem Fox 9 SO A B 33 eR 45 7 2 1 Implementation of the subgoal chain display applet 47 To Lie server prosta oa aei on Rx Zo xod kp Re wok um RO ee ES 48 8 Critical Appraisal 50 9 Testing Summary 52 Oa The cu a ee De BRAD R EM cc 53 10 User Manual 54 10 1 The JSAND Applieation 4 oo RR ER ee 54 10 1 1 The Wintermute display mode 0 56 10 1 2 The Island display mode llle 56 10 1 3 The Jsand display mode gt o e so a saasaa osaami aea 56 10 2 The JSAND Applet 2 lh 58 10 3 The Subgoals Applet oaoa les 58 10 4 The Server Program 2 ea o RR eb eee Re EG Rn 59 11 Maintenance Document 60 11 1 Compilation of the source files cles 60 11 2 Alteration of the parsers for different training programs 61 11 3 Testing strategies used o oo 61 i CONTENTS CONTENTS 12 Software Listings 64 12 1 The Jsand Application Jsand java 22e 64 12 2 The JSAND applet JsandApplet java 74 12 3 The neural network storage module ANN java 82 12
6. MAXX MINX zoom g setColor Color red int counter if boundaries length gt 3 for counter 0 counter boundaries length 3 counter 4 g drawLine width 2 int boundaries counter scale height 2 int boundaries counter 1 xscale width 2 int boundaries counter 2 scale height 2 int boundaries counter 3 xscale private void draw a boundary Graphics g int boundary double scale height MAXX MINX zoom g setColor Color red int counter num 0 if boundaries length gt 3 for counter 0 counter boundaries length 3 counter 4 if num boundary g drawLine width 2 int boundaries counter scale height 2 int boundaries counter 1 scale width 2 4 int boundaries counter 2 scale height 2 int boundaries counter4 3 scale num 103 12 5 The input output display module I0SpaceCanvas java Software Listings private void draw axes Graphics g double inc double scale height MAXX MINX zoom int counter g setColor Color black g drawLine width 2 int MINX xscale height 2 width 2 int MAXX xscale height 2 g drawLine width 2 height 2 int MINY scale width 2 height 2 int MAXY scale double pos 0 0 while pos lt MAXX pos inc g drawLine width 2 int posxscale height 2 2 width 2 int pos scal
7. and boundary lines Once generated regions polygons hyperbands and boundary lines are passed to the IOSpaceCanvas object for display 2 hBand java is a simple class definition which is used by both the I0SpaceCanvas class and the ANN class When instantiated as an object it represents a single hyperband 3 IOSpaceCanvas java contains all the code used to plot an IO space display on to a Java AWT canvas All three display modes Jsand Island and Wintermute are managed by this class and when instantiated the chosen display mode and attributes may be altered by toggling public flags in the IOSpaceCanvas object 7 2 The subgoal chain display The subgoal chain display applet see Figure 7 7 allows the changing behaviour of a neural network to be plotted in output space Since output space is n dimensional the applet must reduce each n dimensional point to a two dimensional point as described in the project plan The two dimensional display contains each n dimensional subgoal represented by coloured Ah 7 2 The subgoal chain display Project Details Data from the ANN simulator Try to connect to host orkney dcs Server initialised Output 0 55098301 17225647 0 8600180149078369 o orkney dcs SubGoal 1 0 5633640289306641 0 8563230037689209 Goal 0 7699999809265137 0 7699999809265137 0 76 SubGoal 0 0 5509830117225647 0 8600180149078369 SubGoal 1 0 5633640289306641 0 8563230037689209 Start
8. double 0 grad y2 yl x2 x1 ix yl grad x1 hbc hbm grad iy hbm x ix hbc System out println horizontal ix ix iy Hiy else grad x2 x1 y2 y1 ix xlcgrad hbc grad yl 1 gradxhbm iy hbm x ix hbc System out println vertical ix ix iy iy Weonly deal with two intersections of the polygon so either start subregion b or close it for good if subregionB size 0 intersect intersect intersect subregionA addElement new Double ix subregionA addElement new Double iy ix iy ix subregionB addElement new Double ix subregionB addElement new Double ly 02 12 3 The neural network storage module ANN java Software Listings Addthe x and y vertices of the current edge to subregion a or b if lintersect 1 subregionA addElement new Double x2 subregionA addElement new Double y2 else subregionB addElement new Double x2 subregionB addElement new Double y2 j j If two sub regions were generated from this region remove the origional region from the vector and add the two sub regions if subregionB size gt 0 amp amp lintersect 1 result addElement rVecToArray subregionA result addElement rVecToArray subregionB else result addElement region j return
9. g fillRect 0 0 current size width current size height Then plot the canvas according to the selected boolean switches and available data if no plot data is generated nothing can be plotted if threed amp amp translated null for counter 0 counter translated length counter if counter stnum step draw plane double translated counter g true else draw plane double translated counter g false else if step amp amp regions null 1 region centre ANN regionCentre double regions stnum 08 12 5 The input output display module I0SpaceCanvas java Software Listings draw region double regions stnum region_centre g draw_a_boundary g stnum mark point region centre 0 region centre 1 8 g label new String stnum 1 region centre 0 region centre l g else if draw_regions amp amp regions null 1 for counter 0 counter lt regions length counter region_centre ANN regionCentre double regions counter draw region double regions counter region centre g if number regions amp amp region activation label new String counter 1 region centre 0 region centre l g if draw bands amp amp bands null amp amp threed for counter 0 counter bands length counter draw hyperband bands counter g if
10. initial int xoffset 50 Border width at either side of display int bottomy 0 space at bottom of screen for text int yoffset current size height bottomy 2 int width current size width 2 xoffset int link oldx oldy Set up a graphics buffer to draw the new image into if offlmage null offlmage createlmage current size width current size height if offlmage getWidth this current size width offlmage getHeight this current size height offlmage createlmage current size width current size height 112 12 6 The subgoal display applet SubgoalApplet java Software Listings If the flag is set to perform double buffering then draw into the buffer if dbuffer g offlmage getGraphics else g window Block out the background in white first g setColor Color white g fillRect 0 0 current size width current size height If the user has selected marking of subgoals then draw them too in the appropriate positions on screen if goals nul f g setColor Color green g drawLine goals 0 x xoffset goals 0 y yoffset goals goals length 1 x xoffset goals goals length 1 y yoffset g setColor Color blue label_subgoal initial goals 0 x xoffset goals 0 y yoffset 8 g label subgoal goal goals goals length 1 x xoffset goals goals length 1 y yoffset 8 g if
11. printf Error on socket bind n exit EXIT_FAILURE return sock Initialises a socket that has allready been created to a specific network address void init_sockaddr struct sockaddr_in name const char hostname unsigned short int port struct hostent xhostinfo name sin family AF INET name sin port htons port hostinfo gethostbyname hostname if hostinfo NULL fprintf stderr Unknown host s n hostname exit EXIT_FAILURE name gt sin_addr struct in_addr hostinfo gt h_addr 124 Chapter 13 Status Report At the end of the implementation phase most software functions described in the initial plan along with the enhancements described in the revised timetable chapter have been completed Both the JSAND and Subgoal applets are able to operate correctly in their client server configurations displaying data in real time sourced from training programs provided by the Dr Weir s research group The network editor function of the JSAND application build has not yet been completed however since JSAND may import textual descriptions of a neural network this function is not critical for the end user to be able to make full use of the JSAND software Instead of editing weight values of neural networks within JSAND using slider bars as originally envisaged users can edit textual description files of neural networks instead JSAND does allow the n
12. wcount Weight w Weight n input Weights elementAt wcount pout println FROM numberNeuron w from WGHT w weight pout println FROM Inputs length Hidden length 1 4 WGHT n biasWeight j Output neuron pout println NURN 4 numberNeuron Output pout println NTYP OUTPUTNEURON for wcount 0 wcount Output inputWeights size wcount Weight w Weight Output inputWeights elementAt wcount pout println FROM numberNeuron w from WGHT w weight 3 j pout println FROM Inputs length Hidden length 1 WGHT FOutput biasWeight Bias neuron pout println NURN Inputs length Hidden length 1 pout println NTYP BIASNEURON for ncount 0 ncount lt Hidden length ncount pout println TO ncount Inputs length WGHT Hidden ncount biasWeight pout println TO Inputs length4 Hidden length WGHT Output biasWeight fout close catch Exception e System out println File error on save e Give a neuron object in the 2 n 1 network a unique integer identifier private int numberNeuron Neuron n int counter result 1 for counter 0 counter lt Inputs length counter if n equals Inputs counter result counter if result lt 0 for counter 0
13. 25 5 2 Software modules required Implementation and Testing Plans a two dimensional floating point array This array will contain data displayable as either a colour representation of the network I O mapping or projected as a three dimensional altitude map of the network I O mapping The display module will also receive control information from the network storage module informing it that the array should be recalculated as the network weight state has just been changed If a specific part of the network is being targeted for display a subsection of the network which has the topology 2 n 1 then information will also be passed to the hyperband display module informing it which part of the network is to be displayed I O Space rendering module This renders the contents of the array to an XGL window The three dimensional display module is used to generate appropriately coloured polygon data for final output to the screen All output is done via XGL library calls Polygon generator for 3d display This module given the complete 2d array of current I O mapping will generate a database of 3d polygons for later rendering via XGL The decision as to whether this database should be regenerated for each repaint of the XGL window or only when the network data is to be changed will be taken in the light of performance data taken from a working implementation 5 2 3 Overview of subgoal chain display modules Subgoal chain GU
14. The Step Forward 56 10 1 The JSAND Application User Manual 23J SAND Java Simple Artificial Neural Display Draw Hyperbands Display Axes C Number Regions Show Boundaries 7 Step Regions Hidden Units end Figure 10 4 An Island mode display of two hyperbands and a boundary line Step Back and Step Regions buttons may be used in any of the three display modes to show only one selected region However in Jsand mode all regions are still drawn in wire frame with the selected region appearing solid Screen updates when dragging a Jsand mode display are faster when the Step Regions option is selected to enable wire frame plotting SETTETETCTTEITTITTTTM V Draw Hyperbands Display Axes Number Regions Show Boundaries Step Regions Hidden Units Figure 10 5 An jsand mode display of two hyperbands UB ae 10 2 The JSAND Applet User Manual 10 2 The JSAND Applet The JSAND applet is operated in the same way as the JSAND application The applet has no file menu instead having a Connect to Server button in the lower part of the window along with a writeable text box where the user can enter a server host name Once the name has been entered and Connect to Server pressed the applet will begin streaming training data from the specified server If no server was available an error message will be printed in the Java console and no streaming will take place 10 3 The
15. amp start point null Display angle and magnitudes are calculated according to the magnitude between initial and final output states of the network the initial state and the goal are then 116 12 6 The subgoal display applet SubgoalApplet java Software Listings assumed to be at opposite ends of the display window initial on the left and goal on the right side of the display display scale double display width ndPoint magnitude goal point position start point position display_angle ndPoint angle goal point position start point position Generate a result array for xy coordinate pairs to be displayed result new xyPoint nd outputs size int counter 0 while counter lt result length result counter ndPoint nd_outputs element At counter xy start_point goal_point display scale display angle display_width counter j j return result public xyPoint goals 1 double magnitude angle display_scale display_angle xyPoint result null if goal point null amp amp start point null Display angle and magnitudes are calculated according to the magnitude between initial and final output states of the network the initial state and the goal are then assumed to be at opposite ends of the display window initial on the left and goal on the right side of the display di
16. center new double 2 Convert each 2d region into a to 3d plane with depth index It does this by adding a z coordinate for each xy pair The z is the y from the 2d region and the new y coordinate is the activation value height for the now xz point in the network This makes the initial translation of the planes easy as z is depth into the screen The first element of the 3d plane array will be used later when sorting the planes for screen depth allowing hidden planes to be painted over at plot time for counter 0 counter subregions length counter region double subregions counter The plane adding space for z coordinates and z depth sort value 89 12 3 The neural network storage module ANN java Software Listings plane new double region length 2 3 point 0 for vertice 0 vertice lt plane length vertice if vertice 3 1 center 0 region point 1 center region point plane vertice regionActivation center else plane vertice region point result counter Object plane return result public double regionActivation double xy double x weight y weight bias weight excitation 0 0 double out Output biasWeight int h for h 0 h Hidden length h x_weight Weight Hidden h inputWeights element At 0 weight y weight Weight Hidden h inputWeights e
17. counter lt Hidden length counter 85 12 3 The neural network storage module ANN java Software Listings if n equals Hidden counter result counter Inputs length if result lt 0 amp amp n equals Output result Inputs length Hidden length return result public double calculateBoundaries Object regions Vector lines new Vector int rgn h vertex double x1 y1 x2 y2 ix iy grad x weight y_weight excitation 0 0 boolean plabove p2above For each region calculate a boundary line for rgn 0 rgn lt regions length rgn int intersect 0 double i new double 4 double xterm 0 0 double yterm 0 0 double cterm Output biasWeight double region double regions rgn double xy regionCentre region for h 0 h Hidden length h x_weight Weight Hidden h inputWeights element At 0 weight y weight Weight Hidden h inputWeights element At 1 weight excitation xy 0 x weight xy l y weight Hidden h biasWeight if excitation gt kconstant amp amp excitation lt kconstant xterm Weight Hidden h outputWeights elementAt 0 weight x_weight kconstant 2 0 yterm Weight Hidden h outputWeights elementAt 0 weight y weight kconstant 2 0 cterm Weight Hidden h outputWeights element At 0 weight Hidden h biasWeight kconstant 2 0 0
18. else Data arriving on an already connected socket if read from applet i lt 0 close i FD_CLR i amp active fd set Forks off an ANN simulator process with inputs given across the web int fork simulator int filedes 1 int pid nbytes nbytes read filedes buffer MAXMSG If this process is the newly forked simulator process act on that if pid fork 0 dup2 www _chan_in 0 fileno stdin Attach www chan in to stdin printf Server thread starts here write filedes testing 7 execl SERVER SERVER char 0 printf The Server is exiting for this process i n pid dup2 filedes fileno stdout Attach www chan out to stdout Call the main function of the simulator to start the ball rolling simulator initialise exit EXIT FAILURE j printf PID i continues Wn pid 123 12 7 The server program SimServ c Software Listings Create a socket to the desiered port on this machine int make socket unsigned short int port int sock struct sockaddr in name Create the socket here sock socket PF INET SOCK STREAM 0 if sock lt 0 printf socket creation error n exit EXIT FAILURE Give the socket a name name sin_family AF_INET name sin_port htons port name sin_addr s_addr htonl INADDR_ANY if bind sock struct sockaddr amp name sizeof name lt 0
19. result v mag vectorA v mag vectorB Generates the 2d representatin of this object in the given set of dimensions set true if a dimension is included in the nd 2d calculation public xyPoint xy ndPoint initial ndPoint goal double display_scale double display angle int display width xyPoint result new xyPoint 119 12 7 The server program SimServ c Software Listings double magnitude magnitude initial position position Calculate the reference vector int counter double reference new double initial position length for counter 0 counter lt reference length counter reference counter goal position counter initial position counter double current new double nitial position length for counter 0 counter lt reference length counter current counter position counter initial position counter double angle angle initial position position display angle From the input data calculate cartesian coordiantes for the result data result x int magnitude x Math cos angle display scale result y int magnitude Math sin angle display scale return result 12 7 The server program SimServ c Neural Network Simulator Server Program C Version Connecting directly to stubs This Server allows the output from the C based neural network simulator to be forwar
20. temp val counter display textbox appendText temp n display data setGoal new ndPoint val The Done statement is found in the input stream else if temp startsWith DONE display textbox appendText Finished training after iterations iterations finished true try inputStream close catch IOException e 110 12 6 The subgoal display applet SubgoalApplet java Software Listings Unhook the socket connection between the display and the server display data update display disconnect Given a bracketed list of comma separated floating point numbers in a string this function returns an array containing the floats private double parseFloats String flist Strip the brackets around the set of floating point numbers if flist startsWith flist flist substring 0 flist length if flist endsWith flist flist substring 1 flist length 1 Count how many numbers by separating commas are within the text String number int counter comma 1 for counter 0 counter lt flist length counter if flist substring counter counter 1 equals comma t double result new double comma Thenextract the floating point values from the text list comma 0 int value 0 for counter 0 counter lt flist length counter if flis
21. two dimensional input the region is between the two parallel lines that form the hyperband 4 excited state Hyperband activation function Output f x k k Excitation x Region in which the hyperband is calculated Figure 2 11 The semi linear hyperband activation function The semi linear hyperband activation function above f x has three distinct output re gions for a given neuron depending on the input excitation x The activity of a neuron as calculated by the hyperband activation function diagram previous page would be as follows 13 2 8 Approximating I O space using Hyperbands Context Survey 1 f x 0 where x lt k 2 f x x 2k 0 5 where k lt x lt k 3 fx 21wherek x The hyperband associated with each non input neuron can be geometrically represented as part of the I O mapping in I O space as follows with a two dimensional input Input Neuron B Input Neuron A Figure 2 12 A Hyperband drawn in 2d I O space The lines bounding the hyperband correspond to the points at k and k on the hyperband activation function The values of k are chosen to ensure that the semi linear hyperband function gives the best possible approximation of the linear sigmoidal function The position and angle of these lines within output space can be calculated as a function of all a neurons weights and present inputs as described by ISLAND 4 Only binary output values can be
22. 4 The hyperband data storage module hBand java 98 12 5 The input output display module I0SpaceCanvas java 98 12 6 The subgoal display applet SubgoalApplet java 107 12 7 The server program BISSBfV lt e 52 323 ee ee aG 122 13 Status Report 127 ili 2 1 2 2 2 3 2 4 2 5 2 6 244 2 8 2 9 2 10 2 11 2 12 di 5 1 Tol 1 2 T 3 7 4 7 5 7 6 tal 10 1 10 2 10 3 10 4 10 5 10 6 List of Figures A neuron processing UNI eene Remo row Mo RR a QUE EUR RE RR An example feed forward strictly layered neural network A neurons input excitation xo does ey lod geo xm eR 3o 9 XAR exeo dr do A step activation function o e sonus rcp R09 A co s xoPe3 ek p A sigmoidal activation function 2 2 Changes in weight space during search for a goal weight state How the resulting 2d data will be plotted on a plane in the case of linear SAHO e ob ae ee RR a SR kh eae eA ER bee RR Low resolution 10 10 representation of an example network I O space The semi linear hyperband activation function A Hyperband drawn in 2d I O space ro eee es Example of subgoal chain visualisation based on a java user interface Overview of the software modules to be implemented as part of this project The same IO space viewed from the three JSAND display modes The Wintermute display mode with ten hyperbands Splitting the display into regions
23. 5 if excitation gt kconstant cterm Weight Hidden h output Weights element At 0 weight double m xterm yterm double c cterm yterm 86 12 3 The neural network storage module ANN java Software Listings for vertex 0 vertex lt region length 1 vertex 2 xl region vertex yl region vertex l if vertex gt region length 2 x2 region 0 y2 region l else x2 region vertex 2 y2 region vertex 3 If we detect an intersection begin end splitting the polygon plabove mx xl c lt yl p2above m x2 c lt y2 if plabove p2above m x2 c y2 amp amp mexl c yl if x1 x2 amp amp y1 y2 System out println Error line of zero length else If represent gradient for both vertical and horizontal lines if Math abs y2 yl lt Math abs x2 x1 y2 y1 double 0 grad y2 yl x2 x1 ix yl grad x x1 c m grad ly m ix c else grad x2 x1 y2 yl ix xl grad c grad yl 1 grad m iy m x ix c ifintersect ix ilintersect iy j if intersect 0 1 lines addElement new Double lines addElement new Double lines addElement new Double lines addElement new Double else if intersect 4 for vertex 0 vertex lt intersect vertex 87 12 3
24. Disconnect a m Ee LA Enter the server name here 4 j M initial goal Figure 7 7 An example subgoals applet display circles with a set of line segments superimposed Each line segment represents a change in output state of the network as supervised learning progresses Together these line segments give a 2d representation of the learning path of the network as it is trained towards each subgoal The user is then able to visually determine how far from any subgoal the network is at any point in its training As with the JSAND applet program the subgoal applet receives training data from a remote server program The same server program is used which again simply forwards the ASCII output of the training program to the applet once a network connection has been established The applet then parses this data stream in order to find new output states for display The training program being used was supplied by Dr Weir s research group and makes use of a particular subgoal training algorithm This training program outputs the following collection of ASCII tags e SG Indicates a new subgoal is to follow e 0g O0 Indicates a new goal state is to follow e wc Indicates a new weight state to follow e 0 c0 Indicates a new output state from the network is to follow 46 7 2 The subgoal chain display Project Details Each of these textual tags is followed by the appropriate list of comma separated re
25. Figure 7 3 shows how a single rectangular display region is split into eighteen regions wherever hyperbands intersect The right hand side of the figure shows the resulting regions in an exploded view where regions have been spaced apart from each other to better illustrate the result of the splitting process Once all regions are known the behaviour of the output neuron within that region may be calculated From the mathematical analysis provided by Storer 4 a further straight line formula for each region may be computed This new line is known as the boundary line and represents the line at which the network s output neuron will have an excitation of zero and activation of 0 5 Boundary lines are calculated with respect to each region and are only valid if they intersect the region for which they were calculated Any boundary lines falling outwith their region are discarded from the display Some of the regions shown in the 38 7 1 The JSAND input output space display Project Details Figure 7 3 Splitting the display into regions right hand half of figure 7 3 have boundary lines through them lines which are contained entirely within the separated convex region shapes After computation all boundary lines that were found within regions may be plotted In Storer s original implementation of ISLAND these boundary lines are then used to further split regions into subregions Each subregion was then coloured by fee
26. JOE REGERE k k k k RE ee pe E p eee E X e k ak e kk ak k e k k ie Class ANNReceiver x x Purpose Manages all I O between Applet and ANN Simulator which is running remotely Communication via BSD sockets This class executes as a standalone thread in paralell with the rest of the applet PK ak k ak k k k ak K K k k k ak K K k ak K kk k k ak K k k ak K k K k K ak K k k ak K k k k K k class ANNReceiver implements Runnable public Thread receiveThread This thread which must be active to recieve bytes private DatalnputStream inputStream Holds the input byte stream from reciever public TextField units public IOSpaceCanvas display public JsandApplet applet public ANNReceiver IOSpaceCanvas thedisplay DataInputStream input TextField numhidden JsandApplet theapplet Copy the instance of the reciever object and set up buffer variables and flags associated with decoding incoming frames display thedisplay inputStream input units numhidden applet theapplet receiveThread new Thread this receiveThread start Entry point for the thread on start of execution public void run try Thread sleep 1000 catch InterruptedException e Try to load a description of a neural network from the stream Vector description null description display network loadNetwork inputStream while description null 1 display network generateNetwork des
27. Neuron max 1 int input 1 int hidden 1 int neuron 1 Generate all the required neuron objects then connect them with weight objects and values parsed from the description file int counter for counter 0 counter lt newNeurons length counter newNeurons counter new Neuron for counter 0 counter lt description size counter String temp String description elementAt counter if temp null if temp startsWith NURN temp temp substring NURN length temp length neuron Integer parseInt temp trim 10 else if temp startsWith NTYP temp temp substring NTYP length temp length int ntype Integer parseInt temp trim 10 if ntype INPUTNEURON Inputs input newNeurons neuron else if ntype OUTPUTNEURON Output newNeurons neuron else if ntype HIDDENNEURON Hidden hidden newNeurons neuron 83 12 3 The neural network storage module ANN java Software Listings else if temp startsWith FROM temp temp substring FROM length temp length trim int btwn Integer parseInt temp substring 0 temp indexOf trim 10 temp temp substring temp indexOf WGHT WGHT length temp length trim double weight findDouble temp if neuron biasNeuron amp amp btwn biasNeuron newNeurons btwn addWeight newNeurons neuron weigh
28. RO EORR 17 3 2 1 The use of host operating system libraries 17 2 2 2 Theben ioo uoce mex RE oso we ea a eb WOW RU E dd 17 23 9 Documenbatl H 222 4 4 c 309 GG ede UR LO Xe E Xs Odo E O3 18 3 2 4 Performance constraints oe a en 18 2 29 5 Error handling L4 2224 xou eh heh pu de Xxx 18 3 2 0 Possible enhancements 2222s llle 19 Commentary 20 4 1 Evolution of the project plan e saoe so e mao on a a a D e aaa 20 Implementation and Testing Plans 22 5 1 Software resources to be used during development 22 Dd Whe XGL RDA e ae prei od x ai ea ok ETE ee ok Rr ee 22 5 1 2 The use of the Java JVM and AWT ls 22 CONTENTS CONTENTS 5 1 3 Integrating Java and C software modules via the JNI 23 5 2 Software modules required 22 2 24 5 2 1 Summary descriptions of core modules ls 24 5 2 2 Overview of I O space display modules 25 5 2 8 Overview of subgoal chain display modules 26 5 8 Integration and testing of the modules 0 0 00 0020 27 54 Deliverables usos e ee es eee RA ee ed e e e np ee ee oh 28 6 Revised Project Plan Timetable 31 6 1 Specification changes proposed during February 2000 32 6 2 Revised software development plan llle 33 7 Project Details 36 7 1 The JSAND input output space display 2 36 7 1 1 The Wintermute display mode
29. The neural network storage module ANN java Software Listings lines addElement new Double i vertex else System out println Error wrong number of intersects j j return double rVecToA rray lines public Object calculateRegions Initialise the first region on the list to the dimensions of the screen an array of xy pairs Vector regions new Vector double screen new double 8 screen 0 IOSpaceCanvas MINX screen 1 IOSpaceCanvas MINY screen 2 IOSpaceCanvas MINX screen 3 IOSpaceCanvas MA XY screen 4 IOSpaceCanvas MAXX screen 5 IOSpaceCanvas MA XY screen 6 IOSpaceCanvas MAXX screen 7 IOSpaceCanvas MINY regions addElement Object screen Then split any region in the regions vector regions in two whenever it is intersected by a hyperband edge hBand bands calculateHyperbands int x for x 0 x lt bandslength x For each hyperband edge regions splitRegions regions bands x gradient bands x top regions splitRegions regions bands x gradient bands x bot Object result new Object regions size regions copyInto result return result public Object calculateWintermute Initialise the first region on the list to the dimensions of the screen an array of xy pairs Vector regions new Vector int xpos int ypos 88 12
30. a zoom factor calulated by the actual display virtual display width width zoom double yl y2 scale height MAXX MINX xzoom double x1 MINX 3 Taking x and y coordinates at the edges of the double x2 MAXX 3 virtual display Compute two y coordinates on the centre line of the hyperband yl xl hyperband gradient y2 x2 hyperband gradient Draw the top and bottom border lines of the hyperband on the display with the display center being the origin of I O space g setColor Color blue g drawLine width 2 4 int xl scale height 2 int y1 hyperband top scale width 2 int x2 scale height 2 int y2 hyperband top xscale g drawLine width 2 int xlxscale height 2 int yl hyperband bot scale width 2 int x2 scale height 2 int y2 hyperband bot scale N m m private void draw region double region double region centre Graphics g 100 12 5 The input output display module I0SpaceCanvas java Software Listings double scale height MAXX MINX zoom Polygon p new Polygon int counter for counter 0 counter lt region length 1 counter 2 p addPoint width 2 int region counter scale height 2 int region counter 1 scale if region_activation int rgb int network regionActivation region_centre 400 if r
31. appropriate learning rule During super vised learning the target output of the network and the networks actual output response to the training input can be computed An error value for an output unit can be computed by subtracting the target output of that neuron from the actual output and squaring the result Error value Target output of neuron Actual output of neuron Back propagation learning specifies that the order of the weight changes will be last to first That is weights along links feeding into output neurons will be altered first then those in the last hidden layer and so on with weights from inputs being changed last In this way back propagation is able to pass delta errors backward through the network where each depends on the previous delta errors The weights associated with each neuron can then be altered using that neuron s delta error uu 2 4 Subgoal chains Context Survey value and a suitable learning rule In order to learn successfully the supervised output of the network should approximate the target output as closely as possible that is the value of the output neuron s error should be made as low as possible during training Once this process is complete the network should have successfully have learned a new input output mapping based on the training set 2 4 Subgoal chains An ideal I O mapping is that mapping associated with the global minimum produced by the goal weight state This goal we
32. be displayed in Jsand mode The file menu is selected from the menu bar 1 Figure 10 2 illustrates this menu and the four options provided The user can load 1 a new network using a platform specific dialogue box save 3 a network edit 2 a network and quit 4 the JSAND program The edit network function produces a dialogue box which allows specific neuron weight values to be changed 4J SAND Java Simple Artificial Neural Display File Load New Network Ctrl O Number Regions Figure 10 2 The JSAND application file menu 55 10 1 The JSAND Application User Manual 10 1 1 The Wintermute display mode Figure 10 3 shows JSAND in Wintermute display mode with the boundary line option 4 turned on The boundary line 1 can be clearly seen dividing the lighter left 3 hand side of IO space from the darker right 2 hand side S J SAND Java Simple Artificial Neural Display Draw Hyperbands Display Axes Number Regions Show Boundaries Step Regions Figure 10 3 The Wintermute display mode 10 1 2 The Island display mode Figure 10 4 shows an Island mode display of two hyperbands marked 1 and 2 and a boundary line 3 10 1 3 The Jsand display mode Figure 10 5 shows the same two hyperbands from figure 10 4 in Jsand mode again marked 1 and 2 This Jsand mode display has been rotated by the user dragging the mouse across the IO space display As the mouse moves the display should rotate
33. compilation of the JsandApplet java source along with the three sources below it generates the applet The Java compiler program will automatically compile sources for modules referenced in the current source as long as they are contained in the current working directory At run time the Java virtual machine will then dynamically link 44 7 2 The subgoal chain display Project Details each class file to form either a working JSAND applet or application The three source files common to both applet and application builds provide the following functionality 1 ANN java contains the data storage module described in the plan spread through three classes ANN Neuron and Weight ANN is the central class which contains constructors and methods able to generate a new random 2 n 1 neural network or create a new neural network from an ASCII description in the format described previously The ANN class accepts ASCII descriptions for parsing from a Java input stream This allows neural network data to be streamed from a remote server or loaded from a local file ANN also contains methods to output the current network state to a Java output stream allowing data to be saved to a local file The Neuron and Weight classes are used by the ANN class to represent the data struc ture of a neural network in an object based manner The ANN class also contains meth ods to generate 2d regions 3d polygons lists of hyperbands each an hBand object
34. data storage object At the start we are not connected to the server and this will not change until the user clicks the start button chainDisplay new SubGoalCanvas add chainDisplay data new SubGoalChain 680 chainDisplay validate j public void paint Graphics g 1 If not yet done put all the required components in their places within the applet window if llaidOut Insets insets insets ext reshape 10 insets left 17 insets top 200 13 extension reshape 10 insets left 30 insets top 200 42 go reshape 220 insets left 30 insets top 80 42 106 12 6 The subgoal display applet SubgoalApplet java Software Listings disconnect reshape 320 insets left 30 insets top 80 40 stat reshape 10 insets left 97 insets top 300 13 textbox reshape 10 insets left 110 insets top 400 150 chainDisplay reshape 10 insets left 280 insets top 780 300 laidOut true y j public boolean action Event e Object arg if e target disconnect Disconnect from the server here disconnect return true else if e target go if receiver null receiver read Thread stop if s null disconnect j String host extension getText textbox appendText Try to connect to host host n Attempt to connect to the given server name and then involk the ANN simulator try
35. input output space display JSAND Java Simple Artificial Neural Display the input output IO space display system allows the user to supply a feedforward single layer neural network constrained to a 2 n 1 topology When a neural network weight state has been loaded into JSAND the complete IO mapping of that network will be displayed for that networks current weight state The user has a choice of three possible display modes Island two dimensional Wintermute two dimensional and Jsand three dimensional Figure 7 1 illustrates the same neural network IO space for a network with two hidden units n 2 in Island mode left Wintermute mode 36 7 1 The JSAND input output space display Project Details Figure 7 1 The same IO space viewed from the three JSAND display modes center and Jsand mode right The first two modes Wintermute and Island are so named because of their similarity to previously available IO space display systems As a two dimensional display each pixel at integer input coordinates X Y on the display plane should represent a possible output activation of the ANN The pixels across the X dimension of the display are used to represent the range of possible input values for the first input neuron and the Y dimension the second input 7 1 1 The Wintermute display mode Figure 7 2 The Wintermute display mode with ten hyperbands The Wintermute display mode splits the on screen display area i
36. k ak k k ak SEE ak k k ak K k ACE JSand Display Module x By Duncan McPherson x x The display canvas module is stored in this source x and is common to the applet and java application K ESI k ak obe ak k SERE OR OE EGRE K k k ak K k ak ORE ak Se k k ak K k ak K k K Re import java awt event import java awt import java io import java util import java awt image 06 12 5 The input output display module I0SpaceCanvas java Software Listings import java lang Math KR SEGSESE dE EERE SR OE AEN SEE RARE RR k k k k ak k Eo ak k k k eel x Class IOSpaceCanvas x x Purpose Draws the Input Output space of a neural network Ne Eee GE k ak k oe k k k ak k K k ak K ER k ak K k k ak K k ak ate dee e p k k k K k class IOSpaceCanvas extends Canvas hBand bands null double boundaries null Object regions null Object planes null Object translated null ANN network double zoom 1 scaling of the IOSpace canvas public static final double MINX 25 public static final double MINY 25 public static final double MAXX 25 public static final double MAXY 25 public static final double MAXZ 20 Boolean properties of the display that may be turned on or off public boolean draw_bands false public boolean draw regions true public boolean number regions false public boolean draw axes false public boolean boundary lines false p
37. of the dialog box Panel panel The panel containing the text IOSpaceCanvas display TextField biasval Label biaslabel public EditDialog IOSpaceCanvas being displayed Frame parent frame super parent frame Edit Network Properties false 09 12 1 The Jsand Application Jsand java Software Listings display being displayed setSize 400 225 addWindowListener this panel new Panel panelsetLayout new GridLayout 5 2 biaslabel new Label Bias value panel add biaslabel biasval new TextField 10 biasval setEditable true panel add biasval ok new Button Ok ok addActionListener this close new Button Close close add ActionListener this panel add ok panel add close add panel biasval set Text being displayed network bias setLocation 200 150 show public void actionPerformed ActionEvent event if event getSource close dispose else if event getSource ok display network bias new Double biasval getText doubleValue dispose j j public void windowClosed WindowEvent event dispose public void windowClosing WindowEvent event dispose i public void windowDeiconified WindowEvent event public void windowlIconified WindowEvent event public void windowActivated WindowEvent event public void windowDeact
38. one value of the complete pattern To this end the network must have the same number of inputs as there are values forming the input pattern 1st 2nd Output Layer Layer Layer Output Pattern From Network Tuer Weighted connections V between neurons Neuron processing unit Figure 2 2 An example feed forward strictly layered neural network Each input layer value is then passed forward to every neuron in the next layer These forward connections each have an associated weight value which amplifies or attenuates the value being passed into the neuron The neuron calculates an excitation value from all of its inputs and weights using the following formula 2 3 Supervised Learning Context Survey x Input value from previous neuron Di Wik w Weight being applied Figure 2 3 A neuron s input excitation From this excitation value a neuron unit must decide its current output state This is done by means of an activation function in the simplest case a step function When the excitation value computed exceeds a given threshold the step function outputs a 1 otherwise it produces a 0 excited state Step activation function 0 5 0 non excited state V Activating threshold Figure 2 4 A step activation function The weights associated with each onward connection are analog and provide an arbitrary amplification or attenuation to the value being passed into the neuron By th
39. output on either sides of the boundary and intuitively know by the greyness of the regions which side is above and which is below 0 5 In this way the Island and Wintermute display modes are complimentary GUI controls are provided allowing users to toggle the display of hyperbands and boundary lines in either mode 7 1 3 The Jsand display mode Figure 7 4 The same IO space display in 3d from 3 different viewpoints The unique Jsand three dimensional display mode affords the user a highly intuitive way to visualise the current IO mapping of a neural network In this mode regions are calculated as with Island mode and then given a grey scale colouring by finding the activation of their centers From there each X Y point comprising a region is converted to an X Y Z three dimensional vertex The X and Z coordinates are the original X and Y respectively The new Y coordinate in XYZ is the activation of the network at the original XY position in IO space multiplied by an altitude scale factor All the regions in the display are then transformed by rotating their vertices about the y and x axes so that the IO space display is rotated to the current user selected 3d rotation angle and then plotted onto the screen In the Jsand display mode the user is free to change the angle of rotation by dragging the mouse across the display area As the mouse is dragged the rotation angles about x and y NS 7 1 The JSAND input output space display Pro
40. result Given a vector of Double container objects this returns an a double array object containing primitive double elements private Object rVecToArray Vector region double result new double region size int counter for counter 0 counter lt result length counter result counter Double region elementAt counter doubleValue return Object result Returns an array of hyperband objects one for each hidden layer neuron public hBand calculateHyperbands hBand result new hBand Hidden length int counter for counter 0 counter lt Hidden length counter result counter Hidden counter findHyperBand bias kconstant return result 93 12 3 The neural network storage module ANN java Software Listings public static double regionCentre double region double result new double 2 result 0 0 0 result 1 0 0 int counter points 0 for counter 0 counter region length 1 counter 2 result 0 region counter result 1 region counter 1 points j result 0 result 0 double points result 1 result 1 double points return result j Neuron class used to represent one neuron unit of a strictly layered feed forward artificial neural network A neuron object contains all weights and pointers to the connecting input and output neuron s cla
41. s new Socket host 5555 is new DatalnputStream s getInputStream os new PrintStream s getOutputStream os print ANNSIMSERVER textbox appendText Server initialised n data new SubGoalChain 680 chainDisplay Create a new thread which handles all output from the newly connected server receiver new SubGoalReceiver this is catch Exception ex textbox appendText Unable to connect to host n return false return true return false 107 12 6 The subgoal display applet SubgoalApplet java Software Listings public void disconnect if s null try s close catch Exception ex s null 3X SEE ROE E SE SEE SERRE OR k k k EAE k k k kk k k k k k k ak k k k ak k k k k k k Class SubGoalReceiver Purpose Manages all I O between Applet and ANN Simulator which is running remotely Communication via BSD sockets This class executes as a standalone thread in parallel with the rest of the applet Nee ak k a OE k a K k k ak k ak k E SEHE E XE ak k k k ak K k k k K Se e k ak K k k k K k class SubGoalReceiver implements Runnable private DatalnputStream inputStream Holds the input byte stream from reciever Thread read Thread This thread which must be active to recieve bytes SubgoalsApplet display Tokens in the input stream which are parsed in order
42. showgoals true for link 0 link lt goalslength link 1 mark subgoal goals ink x xoffset goals link y yoffset 8 g label_subgoal new String link goals link x xoffset goals link y yoffset 8 g j else mark subgoal goals 0 x xoffset goals 0 y yoffset 8 g mark subgoal goals goals length 1 x xoffset goals goals length 1 y yoffset 8 g j If a subgoal chain has been sucessfully 113 12 6 The subgoal display applet SubgoalApplet java Software Listings generated then display it g setColor Color red if chain nul oldx 0 oldy 0 for link 0 link lt chain length link g drawLine oldx xoffset oldy yoffset chain link x xoffset chain link y yoffset oldx chain link x oldy chain link y if showattained true for link 0 link lt chainlength link mark point chain link x4 xoffset chain link y yoffset 4 g And plot the entire buffered image to the screen in one go if dbuffer window drawImage offlmage 0 0 this Procedure used to draw a point in the subgoal display private void mark_point int x int y int size Graphics g g setColor Color red g fillOval x int size 2 y int size 2 size size g setColor Color blue g drawOval x int size 2 y int size 2 size size Marks a subgoal item
43. space it is possible to use hyperbands to calculate a semi linear approximation of the neural I O mapping A hyperband is the linear part of a semi linear activation function see figure below This approach computes approximate activation characteristics for each neuron and then combines the information to give a complete I O space display 12 2 8 Approximating I O space using Hyperbands Context Survey A previous project ISLAND 4 took this approach to generate a real time network display allowing the I O space of a network with two floating point inputs to be displayed Island produced a two dimensional display of the binary output of the network For this project an analog display of the network output will be produced rendered as an altitude on a three dimensional repesentation of a networks I O mapping Also networks used within ISLAND were required to only have one hidden layer of neurons a limitation which will be partially alleviated in this implementation by allowing any portion of a network with a corresponding 2 n 1 topology to be used to generate an I O space display Networks with the topology 2 n m may also be catered for as an extension to the project by displaying several output space mappings one for each of m output neurons Computing a neurons activity using a hyperband activation function allows a region to be determined within I O space where that neurons activity characteristics are known For a
44. where n represents an arbitrary number of intermediate hidden layers 15 3 1 Functional Requirements Problem Specification of neurons each of which contain one or more neuron units I O space will be displayed as a three dimensional altitude map projected onto a two dimensional window based display Arbitrary resolutions of height map will be possible to represent the underlying I O space mapping in greater detail An upper bound for display resolution should be made according to available memory and a maximum polygon quota giving acceptable system performance The display window will allow the three dimensional display of I O space to be rotated in real time by the user and both perspective and orthogonal projections will be possible In addition a region of a network may be chosen which corresponds to a 2 n 1 mapping and an I O space display for this region only may be generated 3 1 3 Visualisation of Subgoal chains The visualisation of subgoal chains will be implemented via a two dimensional window based user interface An initial point will be chosen on the left hand side of the window to represent the initial output state of the network The reference vector between this initial output state and the goal output state will be transformed in two dimensions so that the goal output state lies on the right hand side of the window The goal output state can then be plotted in the window and all other output space points
45. 10 val parseFloats temp substring temp indexOf 1 temp length trim int counter temp SubGoal sgnum for counter 0 counter lt vallength counterd 4 if counter vallength 1 temp val counter else temp vallcounter display textbox appendText temp n display data addSubGoal val sgnum Anew output state is found in the input stream else if temp startsWith OUTPUT iterations 109 12 6 The subgoal display applet SubgoalApplet java Software Listings val parseFloats temp substring temp indexOf 1 temp length trim int counter temp Output for counter 0 counter lt vallength counterd 4 if counter vallength 1 temp val counter else temp val counter display textbox appendText temp n display data addOutput new ndPoint val Agoal state is found in the input stream only register this once since the goal will be the same throughout the training period we are displaying else if temp startsWith GOAL amp amp goal false 1 goal true val parseFloats temp substring temp indexOf 1 temp length trim int counter temp Goal for counter 0 counter lt vallength counter 4 if counter vallength 1 temp 4 val counter else
46. 3 The neural network storage module ANN java Software Listings int res 50 double width IOSpaceCanvas MAXX IOSpaceCanvas MINX res double height IOSpaceCanvas MAXY IOSpaceCanvas MINY res System out println Width width height height for xpos 0 xpos lt res xpos for ypos 0 ypos lt res ypos double region new double 8 region IOSpaceCanvas MINY double ypos 1 height region 6 IOSpaceCanvas MINX double xpos 1 width region 7 IOSpaceCanvas MINY double yposxheight regions addElement Object region Object result new Object regions size regions copyInto result return result region 0 IOSpaceCanvas MINX double xpos width region IOSpaceCanvas MINY double yposxheight region 2 IOSpaceCanvas MINX double xpos width region 3 ypos 1 xheight region 4 IOSpaceCanvas MINX double xpos 1 width 5 6 7 IOSpaceCanvas MINY double Generates a list of plane objects Each plane is represented as a double array element 0 Plane depth in Z for depth sorting 7 1 xyz Uf 4 More xyz coordinate pairs as needed public Object calculatePlanes 1 Object subregions calculateRegions Object result new Object subregions length int vertice point counter double plane region
47. Done public final static int INPUTNEURON 1 public final static int OUTPUTNEURON 2 public final static int HIDDENNEURON 3 public final static int BIASNEURON 7 8 12 3 The neural network storage module ANN java Software Listings The neural network file parsar Parses a complete neural network data structure and returns only the strings in the file that pertain to that structure so that a stream can have one or more data structures parsed by multiple calls to parseNetwork public boolean loadedCorrectly true int biasNeuron 0 Three results from the preparsing int max 0 int num_hidden 0 public Vector loadNetwork DatalnputStream is Vector netDescription new Vector int inputParsed 0 Ensure the network is indeed int outputParsed 0 a valid 2 n 1 neural network int neuron 1 num hidden 0 biasNeuron 1 boolean result false boolean foundDone false String temp First ensure that the input stream contains a network and then pass all the relevent lines from the input stream to the generator try temp is readLine catch Exception e temp null while temp null amp amp foundDone 1 temp temp trim if temp startsWith NURN netDescription addElement Object new String temp temp temp substring NURN length temp length neuron Integer parseInt temp tri
48. ELLO if strncmp HELLO buffer strlen HELLO 0 fork_simulator filedes return 0 x lee 12 7 The server program SimServ c Software Listings i int main void extern int make socket unsigned short int port int sock fd set active fd set read fd set int i struct sockaddr in clientname size t size Create the socket and set it up to accept connections sock make socket PORT if listen sock 1 lt 0 printf Error on listen for socket n exit EXIT FAILURE Initialize the set of active sockets FD ZERO amp active fd set FD_SET sock amp active fd set while 1 1 Block until input arrives on one or more active sockets read fd set active fd set if select FD SETSIZE amp read fd set NULL NULL NULL lt 0 printf select Wn exit EXIT FAILURE j Service all the sockets with input pending for i 0 i lt FD SETSIZE i if FD ISSET i amp read fd set if i sock Connection request on original socket int new size sizeof clientname new accept sock struct sockaddr amp clientname amp size if new 0 printf accept n 122 12 7 The server program SimServ c Software Listings exit EXIT FAILURE fprintf stderr Server connect from host 4s port Ahd Wn inet ntoa clientname sin addr ntohs clientname sin port FD SET new amp active fd set
49. I manager module This module handles the storage of previous subgoal positions on the 2d subgoal display and also calculation of the networks current output space subgoal position via class methods provided in the network data storage module Once calculated these position data points are also supplied to the cubic spline generator module which generates an array of points which at the resolution of the output display will appear to be a smooth cubic spline curve This curve data along with the output space points will be supplied to the subgoal chain rendering module which handles output to an AWT window 26 5 3 Integration and testing of the modules Implementation and Testing Plans Cubic spline generation module This module given the current list of output space points to be plotted to the window canvas generates a further array of points which may be used to plot a smooth cubic spline curve The number of points generated by this module will be dependent on the current display resolution with a maximum of one point per vertical pixel of the display window canvas Subgoal chain GUI canvas module This module handles the display of the subgoal chain window and plots the points and associated cubic spline curve as required by the AWT It also handles user interaction with this window in order to change display options associated with subgoal chains 5 3 Integration and testing of the modules Each modules s funct
50. Internet web pages Additionally instead of using the JNI as a communications bridge between the external training program and the display applet a client server system has been implemented to transfer training data from a training program running remotely to the client applet This has the advantage of making the display client applets completely platform independent as they interact with the remote server program to receive updates A C based server program has been implemented to allow this func tionality Finally the completed input output space display system may also be built in the form of a Java application In addition to the functionality provided by the applet build the application build also allows neural networks to be filed on disc and edited manually The revised development timetable proposed was as follows 93 6 2 Revised software development plan Revised Project Plan Timetable Week 2 7 11th February Discuss this timetable with mkw Possible meeting with mkw s research group members to receive source code for ISLAND and their altered version of the subgoal display component Initial developments in Java using Storers original ISLAND source Alteration of subgoal chain program to allow selection of inputs Initial changes and development to be submitted to mkw on Tuesday 15th February Week 3 14 18th February Feedback on initial changes to subgoal code from mkw Discussion of Hyperband theory and enhancements of Sto
51. Subgoals Applet Applet Viewer SubgoalsApplet class E Applet Enter the server name here orkney do Start Disconnect Data from the ANN simulator d Output 0 7568219900131226 0 9999740123748779 0 gt Output 0 756945013999939 0 9999740123748779 0 7 Output 0 7570670247077942 0 9999740123748779 0 Output 0 7571899890899658 0 9999750256538391 0 Output 0 757311999797821 0 9999750256538391 0 7 Output 0 7574350237846375 0 9999750256538391 0 j Ej J MEE E o Applet started Figure 10 6 The subgoal display applet in a Solaris desktop Once loaded in a web page the subgoals applet see Figure 10 6 may be connected to a server by entering the server name in the server text box and then clicking the Connect button As with the JSAND applet data is streamed from the remote server after this action has been taken 58 10 4 The Server Program User Manual The output path is displayed in the lower part of the applet window On connection to the server an initial network state and a goal network state are plotted and marked on screen From there new subgoal points and output state line segments are added as they are received by the applet The output and goal states received are also printed in a Java text box on the top left of the applet window States in the text box may be cut and pasted into other application programs 10 4 The Serv
52. The real time visualization of artificial neural output during supervised learning Senior Honours Project 14th of April 2000 Duncan McPherson University of St Andrews Abstract This project report describes the design and implemenentation of two visualization tools which graphically show the changing state of an Artificial Neural Network ANN during supervised learning JSAND Java Simple Artificial Neural Display allows the user to plot the input output mapping of an ANN An I O mapping represents the complete set of output patterns of an ANN for all possible combinations of input pattern according to the current weight state of that network JSAND is able to render a three dimensional surface which accurately conveys the current I O mapping of an ANN to the user as it changes dynamically in real time during supervised learning Additionally JSAND is able to provide two dimensional I O mapping plots in a variety of styles The second tool is able to plot a two dimensional path of the changing behaviour of an ANN in output space as supervised learning takes place Both tools are able to be built as Java applet clients In client form these applets acquire ANN training data from a remote server which forwards data from a training program running on that server Additionally the JSAND tool may also be built in Java application form where it allows basic ANN editing functions along with the ability to load and save neural networks
53. a text Jat SERE SOR RE He opp k k k OE k k k kk ak k k k k k o k ak k de x Class SubGoalsApplet Main applet class Purpose The core applet routine which intialises the applet GUI and begins communication with the server program Once communication has begun a separate threaded object SubGoalleceiver is used to allow the GUI to remain responsive whilst receiving new points from the ANN simulator MESE ak ak k k ak K ROUGE RE K k k ak k ak K k op ex ak K k oe Eo public class SubgoalsApplet extends Applet X X X X TextField extension TextArea textbox Button go disconnect SubGoalCanvas chainDisplay SubGoalReceiver receiver SubGoalChain data Label ext stat private boolean laidOut false 105 12 6 The subgoal display applet SubgoalApplet java Software Listings Communications with server variables DatalnputStream is null Socket s null PrintStream os null public void init setFont new Font Helvetica Font PLAIN 14 ext new Label Enter the server name here add ext stat new Label Data from the ANN simulator add stat textbox new TextArea 10 40 add textbox go new Button Start add go disconnect new Button Disconnect add disconnect extension new TextField 20 extension setEditable true extension setText orkney dcs add extension Create a subgoal canvas and
54. aining program It then passes newly parsed data to a SubGoalChain object and requests that the display is re painted so that the user may see new data points appear in real time When instantiated the SubGoalReceiver class runs as a separate thread to ensure that the subgoal display can be updated and the GUI still functions while the SubGoalReceiver thread awaits new data e SubgoalCanvas This module provides all graphical display functions essentially the required Java paint method which redraws the display The data used to produce 4T7 7 3 The server program Project Details the display is held within a SubgoalChain object To speed up redraw operations the SubgoalCanvas object caches lists of xyPoint objects that have been previously computed from ndPoint objects e SubgoalChain This is the core data storage module in the applet program It pro vides methods used to add new subgoals goals and output states to display When instantiated this object must be known to both the current SubGoalReceiver object and SubGoalCanvas object This allows the receiver to add newly parsed data and the canvas object to then display this data when a repaint operation is requested Data points are stored as lists of ndPoint objects e xyPoint Provides a simple container data structure for a pair of integer values e ndPoint Provides an abstract data type for manipulating n dimensional points An ndPoint object may be constructed and th
55. al numbers in ASCII format Lists of real numbers are delimited by curly braces The applet parses this data stream and displays the resulting information on the two dimensional display canvas as it is received Additionally a Java textbox is used to display all elements of the data stream that have been used in creating the display This textbox contains all output states subgoals and goals being displayed which may then be cut and pasted by the user to another application or examined within the scrollable text box One advantage of applet programs is that vector graphics being displayed such as the path of the network may be output as postscript vector graphics The appletviewer program supplied as an integral part of the Java Development Kit JDK allows postscript files to be printed out from the applet window 7 2 1 Implementation of the subgoal chain display applet The subgoal chain display applet is built from a single Java source file SubgoalApplet java which containing the following class definitions e SubgoalApplet This class contains all of the applet GUI and widgets User controls are provided to allow connection to the remote server program All Java events are handled by this class and an initial connection to the server is established by the class before a SubGoalReceiver thread is instantiated to handle the connection e SubGoalReceiver This module contains all code necessary to parse incoming ASCII data from the tr
56. and supplying a client side display Java applet with weight state data via a Berkeley socket connection This form of connection between an applet and 32 6 2 Revised software development plan Revised Project Plan Timetable its originator connection is permitted under the Java security model provided the users web browser has not been configured to prohibit applets making any Internet connections e Development of the input output mapping display component has been decided to continue with the immediate re implementation in Java of parts of Storers original C code implementation of ISLAND Once a complete Java implementation is available this component of the SH project can also be evaluated for use as a web applet and any changes necessary to achieve this end made 6 2 Revised software development plan The proposed changes section of this chapter and the following software development timetable was submitted to Dr Weir on February 4th 2000 From there the seven week schedule was followed reasonably closely though software development work continued up to and including week eight In the light of implementation experience the XGL graphics libraries have not been used for 3d rendering Instead as partially described in the possible enhancements section of the original plan all display code has been written in Java and the finished software systems for both subgoal chain and input output space display are able to run as Java applets in
57. and the path between them will similarly be transformed in two dimensions using the same transformation This will ensure that the two dimensional area where the networks output states are likely to be plotted during training is to be the central area of the window The calculations used to approximate n dimensional output states to two dimensional dis play points are described in section 3 7 With training to each subsequent subgoal output state a new output space point will be plotted in the window In addition to plotting output space points the path between them must be represented also as detailed in section 3 7 to help visualise trends in output state changes The cubic spline curve being plotted will be determined using the two previous two dimensional output space points 16 3 2 Non Functional Requirements Problem Specification IE Visualisation of subgoal chains Output state path of network Output state at subgoal Figure 3 1 Example of subgoal chain visualisation based on a java user interface 3 2 Non Functional Requirements 3 2 1 The use of host operating system libraries The target platform for this project is to be a Solaris based PC computer as found in the honours laboratory Several different software tools will be used to during the development of this software in particular Sun Microsystems JDK Java Development Toolkit and also Sun s XGL graphics libraries for use within the Solar
58. avac Jsand java command e Similarly the applet is built using javac JsandApplet java e The Subgoals applet may be built with javac SubgoalsApplet java 060 11 2 Alteration of the parsers for different training programs Maintenance Document e The Server program is built using gcc with the following command gcc SimServ c Trainer c lm lnsl lsocket where the Trainer c program is the training pro gram with main renamed simulator initialise For the above commands to work it is assumed that all source files are contained in one working directory 11 2 Alteration of the parsers for different training programs The parsers used to read ASCII input from a training program may be altered simply for use with different training programs All tokens used by the parser are held as constants at the top of the class file Should a training program output have a different output syntax to that expected minor changes will be needed to the subgoal or the jsand parser code In the case of the jsand parser the structure of neural network data internally is very flexible Since a neural network module is a Java object which can have elements weights and neurons added to dynamically the parser is able to build a complete network object before discarding the old one and updating the display The ANN class contains methods to add weight and neuron objects after the network has been constructed 11 3 Testing strategies used Durin
59. ble region Vector subregionA subregionB boolean intersect false Vector result new Vector Gothrough each region in turn and decide whether it is intersected in two places by this hyperband If so split the region into two regions for rgn 0 rgn lt regions size rgn if double regions elementAt rgn length gt 0 region double regions elementAt rgn Find out the number of intersections between this hyperband line and the region given Keep intermediate results from this calculation to help split the region subregionA new Vector subregionB new Vector intersect false for vertex 0 vertex lt region ength 1 vertex 2 xl region vertex 91 12 3 The neural network storage module ANN java Software Listings yl region vertex 1 if vertex gt region length 2 x2 region 0 y2 region l else x2 region vertex4 2 y2 region vertex 3 If we detect an intersection begin end splitting the polygon plabove hbmx x1 hbc lt y1 p2above hbmx x2 hbc lt y2 if plabove p2above hbm x2 hbc y2 amp amp hbmxx1 hbc y1 if xl x2 amp amp yl y2 System out printIn Error line of zero length else If represent gradient for both vertical and horizontal lines if Math abs y2 yl Math abs x2 x1 y2 yl
60. boundary lines amp amp boundaries null amp amp threed draw boundaries g if draw_axes amp amp threed draw axes g 5 0 And plot the entire buffered impage to the screen in one go window drawlmage offlmage 0 0 this public void drag int x int y if drag false amp amp threed drag true dragx x dragy y dragxrot xrot dragyrot yrot else if threed xrot dragxrot y dragy 360 if xrot lt 0 xrot 360 yrot dragyrot x dragx 360 if yrot lt 0 yrot 360 translated null repaint j j public void drag finished int x int y 09 12 5 The input output display module I0SpaceCanvas java Software Listings if drag true drag false xrot dragxrot y dragy 96 360 if xrot lt 0 xrot 360 yrot dragyrot x dragx 96 360 if yrot lt 0 yrot 360 translated null repaint j j private void mark point double xp double yp int size Graphics g 1 double scale height MAXX MINX zoom int x width 2 int xpxscale int y height 2 int ypxscale g setColor Color red g fillOval x int size 2 y int size 2 size size g setColor Color blue g drawOval x int size 2 y int size 2 size size private void draw_hyperband hBand hyperband Graphics g Scale the display a constant screen size multiplied by
61. ce opi Polygon Generator rendering module for 3d display ae EA mcm ms ms Rm ARNE Hyperband display Java canvas classes calculation module r and AWT widget controls Network property Z editor module ANN Data storage ANN File loader module 3 and saver module lt Ili Subgoal chain GUI GUI Management manager module of file import export lus Java dialogue box control classes Subgoal chain Cubic spline rendering module lt lt Generator module Java canvas classes and AWT widget controls External training program Figure 5 1 Overview of the software modules to be implemented as part of this project 94 5 2 Software modules required Implementation and Testing Plans The core modules within this project are those which are used by both subgoal chain visualisation and I O space rendering parts of the program These core modules are to be developed in Java Network data storage module T his software module stores the currently loaded neural network within memory It contains java methods used to compute I O characteristics of the network for use by the subgoal chain display modules and the I O space rendering modules Appropriate JNI header files to allow these methods to be accessed from C will be generated for this module Network property editor module T his modules allows the properties of the curre
62. covered in previous meetings There was also further discussion about subgoal chains and possible ways in which they may be visualised 20 4 1 Evolution of the project plan Commentary Week 6 meeting Corrections concerning the neural network theory document were re viewed and several areas of confusion were cleared by MKW Further targets were set for the production of a complete draft project plan Week 7 meeting A mostly complete draft of the project plan was produced and reviewed During this process several technical misunderstandings were discussed Week 8 meeting Final details of the project plan were reviewed and a target was set to produce an overview section between the abstract problem definition and the introduction to neural networks in the context survey A draft of this was to be reviewed during week 8 Week 8 final meeting Several corrections to the overview section were discussed and correction details of the draft project plan were finalised before submission at the end of week 8 21 Chapter 5 Implementation and Testing Plans 5 1 Software resources to be used during development 5 1 1 The XGL library The XGL graphics library will be used for the display of I O space within the solaris X Windows GUI environment This graphics library allows any graphics acceleration hardware available on the host workstation to be used Also via the X windows network graphics model XGL wou
63. cription if display network loadedCorrectly display bands display network calculateHyperbands 79 12 3 The neural network storage module ANN java Software Listings display regions display network calculateRegions display boundaries display network calculateBoundaries display regions if display region activation display regions display network calculateWintermute display translated null display planes display network calculatePlanes units setText new String display network Hidden length display repaint try Thread sleep 500 Half a second delay between frames catch InterruptedException e description display network loadNetwork inputStream else description null display network null j j applet disconnect 12 3 The neural network storage module ANN java PRA AR ee E eR k k k k k k k k ak k oe k k GERE k k ak k k ORE SERE ERE JSand ANN Data Module By Duncan McPherson The ANN data module allows neural networks to be created stored manipulated loaded and saved and also transformed into data useable for display EE SE ae Be k k HE pu op d k ak K OR EE ak K ak k k k ak K k ak K k k de de import java awt event import java awt import java io import java util import java awt image import java lang Math x Class used to represent a 2 n 1 neural network maki
64. d Neural Network FileDialog LOAD dialog show String file dialog getFile if file equals amp amp file equals nu11 if loader null loader loadThread stop loader new ANNLoader display dialog getDirectory file numhidden else if source save Create an ANN with the number of specified hidden units FileDialog dialog new FileDialog this Save Neural Network FileDialog SAVE 05 12 1 The Jsand Application Jsand java Software Listings dialog show String file dialog getFile Some file has been selected and not cancel if file equals amp amp file equals nu11 display network saveNetwork dialog getDirectory file else if source zoomout amp amp display zoom gt 0 25 display zoom display zoom 0 25 display repaint else if source zoomin amp amp display zoom lt 25 1 display zoom display zoom 0 25 display repaint else if source stf if display step 1 display stnum 0 display step true stepon setState true else display stnum if display stnum gt display regions length display stnum 0 j display repaint else if source stb if display step 1 display stnum 0 display step true stepon setState true else display stnum if displaystnum 0 display stnum display re
65. de network code ANN java hBand java IOSpaceCanvas java 1 Data Storage 1 Data Structure 1 Displays IO space on 2 Load Save Stream for hyperbands screen in Island Jsand ASCII network data held in a common Wintermute modes 3 Hyperband calculation class definition 2 3d rotation of polygon from network data data according to mouse 4 Generation of Polygon data drag events as 3d polygons or 2d regions 5 Generation of the boundary line Figure 7 6 The source files comprising JSAND 7 1 6 Implementation of JSAND In most cases each software module in the implementation plan corresponds to a Java class in the final implementation Since JSAND may be built as both an applet and a Java application separate source files are used in order to allow both the applet and application to share common code such as the display canvas module The applet and application builds differ mainly in their initialisation and GUI code Also the application build contains a network editor and GUI functions to load and save a neural network from disc whereas the applet contains client side code required to communicate with a training program running on a remote server Figure 7 6 shows the source tree of JSAND used to build both the applet and application along with numbered descriptions of the software modules contained within each source Compilation of the Jsand java source file and all three sources below it in the tree produce the application and
66. ded to a client applet when that applet involkes the server Duncan McPherson 9 02 00 include lt stdio h gt include lt string h gt include lt sys socket h gt include lt sys systeminfo h gt include lt sys types h gt 120 12 7 The server program SimServ c Software Listings include lt time h gt include lt errno h gt tinclude lt stdlib h gt include lt sys un h gt include lt unistd h gt include lt netinet in h gt tinclude lt netdb h gt Inclusion from network simulator include vector h External function used to initialise the simulator extern void simulator _initialise 7 define SERVER tmp serversocket define MAXMSG 512 define PORT 5555 define ANNSIM SUBGOALAPPLET Welcome sent by the applet define HELLO ANNSIMSERVER Reply sent by the server define SIMSTART SIMSTART Parameterized start request Reads commands from the applet that tries to connect to the port int read_from_applet int filedes char buffer MAXMSG int nbytes nbytes read filedes buffer MAXMSG if nbytes lt 0 Read error creates the following message printf Error on reading from the client n exit EXIT_FAILURE else if nbytes 0 End of file return 1 else Data read fprintf stderr Server got s n buffer printf Hello command is s of length i n HELLO strlen H
67. determined using a hyperband However for the purposes of this project an analog display can be approximated by evaluating the neurons output characterisitics over a range of values in the linear part of the hyperband Once all required hyperbands have been calculated they can be combined to produce a display of the I O mapping for the entire network 14 Chapter 3 Problem Specification 3 1 Functional Requirements 3 1 1 Alteration of network properties The software should provide a GUI based neural network structure and properties editor This should allow changes to the topology of the network and the values associated with each individual neuron The network currently being modified within the editor will be available for direct use by both the I O space display package and subgoal chain display package without the need to save the network Neural networks will be able to be saved as files and loaded from file Compatibility with other neural network file structures will be explored hopefully allowing files created with Wintermute Island SkyNet packages to be imported and exported Also to enable the use of an external training package a specific neural network file will be able to be monitored and reloaded if that file has been changed by another program 3 1 2 Visualisation of the networks I O space The software should be able to render the I O space for any given network with the general topology of 2 n 1
68. ding the center X Y coordinates through the network and finding a resulting output activation Since all subregions generated using this algorithm were convex the center of a subregions was easily calculated by averaging X and Y coordinates from each point defining that subregion The JSAND Island mode display only shows boundary lines no colouring with respect to the output activation of network is performed In Island display mode JSAND colours all regions falling within hyperbands white and regions outwith hyperbands either yellow or orange Yellow regions lie on the on side of one or more hyperbands in the display and orange regions fall on the off side of all hyperbands in the display On and off values are calculated with respect to the output activation of the hidden neuron from which a hyperband is calculated JSAND does however allow users to superimpose red boundary lines over a Wintermute mode display The left hand side of figure 7 2 shows boundary lines clearly separating 39 7 1 The JSAND input output space display Project Details IO space where the network output activation is above and below 0 5 The boundary lines shown in figure 7 2 have been enhanced for clarity with a thinner copy of themselves superimposed in white studying the figure shows how the boundary lines surround the light coloured areas of IO space where the network activation is above 0 5 This allows users to quickly see the analogue
69. e height 2 2 g drawLine width 2 int pos scale height 2 4 2 width 2 int pos scale height 2 2 j pos 0 0 while pos lt MAXY 1 pos inc g drawLine width 2 2 height 2 int posxscale width 2 2 height 2 int pos scale g drawLine width 2 2 height 2 int posxscale width 2 2 height 2 int posxscale Adda text label to a region or area of the display private void label String label double x double y Graphics g double scale height MAXX MINX zoom Font f new Font Helvetica Font BOLD 14 g setFont f g setColor Color blue FontMetrics fm g getFontMetrics f g drawString label width 2 int x scale fm stringWidth label 2 height 2 int yxscale _ 104 12 6 The subgoal display applet SubgoalApplet java Software Listings 12 6 The subgoal display applet SubgoalApplet java JE S oko be qi Ee SEHE GR k S SEO k ERO RE SE k k SE OR k ak k k ak k k k k k k SubGoal Chain Display Applet x By Duncan McPherson 17th of February 1999 Na ak k de ke ak de E SEE SERE EORR k k ak K OE K ak ak ae ae ak Hee ak EROR ORE import java applet Applet import java awt Component import java lang x import java math import java applet import java awt import java awt event import java io import java net import java util import jav
70. e networks learning through n dimensional output space to be plotted in two dimensions 9 2 6 Visualisation of subgoal chains Context Survey 724d d Figure 2 7 a 008 2m Figure 2 8 Though much information will be lost using this approach the two dimensional represen tation should produce an informative display of the networks learning path for different training problems The most important information that it will convey is whether the new output space position after training to a subgoal is getting nearer the actual goal and each subgoal and whether it is heading directly for the goal or not In order to visualise a network s progress towards a known goal output state in two dimen sions a reference vector between the networks initial usually randomly set output state and the goal output state should be found first For the initial output space position an arbitrary point on the two dimensional output space plane can be chosen The reference vector between the networks initial output state and the goal output state can then be plotted with a magnitude calculated as described previously For display purposes the goal output state reference vector will be plotted to the right of the initial output state the appropriate magnitude away All other output states to be displayed will then have their magnitude calculated from the initial output state but will also require an angle to be cal culated This angle
71. e proportionally nearer the final goal state than the previous subgoal However it is known that linear subgoals can still become stuck in local minima under some circumstances The use of non linear subgoal chains may be able to avoid local minima if each subgoal output state is positioned so that training avoids any local minima present in weight space The testing of a new subgoal generation algorithms would benefit greatly from a display of the networks changing output state during training to each subgoal This display should be able to indicate the networks progress or lack of toward each subgoal output state 2 6 Visualisation of subgoal chains Output space is the complete set of possible output states that a network may produce In this respect it has the same number of dimensions as there are output neurons and so a mathematical approximation of n dimensional space must be produced for a two dimensional display A vector calculated between two output states in output space will have both magnitude and an angle which can be calculated in an n dimensional output space as follows Magnitude between two points where dx is a point in dimension x in a space with n dimen sions Angle via Scalar Product between the two vectors a and b using above formula to find magnitudes of each vector The magnitude and angle calculated can then be represented as an output space point on a two dimensional plane This allows the path of th
72. e y D Math sin Math PI 180 double y Then translate all xyz vertice coordinates for counter 0 counter lt planes length counter plane double planes counter region new double plane length 3 2 1 yheight 0 0 point 1 vertice 0 while point lt region length X plane vertice Y planelvertice Z planelvertice yheight Y Scale the altitude of the plane and rotate about x and y axes Y 0 5 Y MAXZ M YxA ZxB N Zx A Yx B Z M Y N M XxC ZxD N Zx C XxD region point N region point Y if depthsort counter lt M depthsort counter 2M And finally the depth index j For colouring this plane average the untranslated y height region 0 yheight double plane length 3 result counter region Finally sort the planes by their depth z coordinate for plotting int i 102 12 5 The input output display module I0SpaceCanvas java Software Listings Object temp double tempindex for counter 1 counter depthsort length counter i counter while i gt 0 amp amp depthsort i lt depthsort i 1 tempindex depthsort i temp result i result i result i 1 depthsort i depthsort i 1 result i 1 temp depthsort i 1 tempindex Poem j return result private void draw_boundaries Graphics g double scale height
73. ed in two dimensions with a start point representing the initial untrained output state of the network and an end point of the ideal goal output state Since output states typically have more than two dimensions a mathematical approxima tion of this path in n dimensional output space is used here Though information will be lost using this approach the two dimensional output path should still be able to convey whether the training is successfully moving the networks output state towards the final goal output state or not Additional plotting of a cubic spline curve between output states in two dimensions should allow output state transitions to be smoothly represented Such a visualisation tool will also allow testing of hypotheses that only certain shapes of subgoal path or sequence are realisable in certain problems Chapter 2 Context Survey 2 1 Introduction to neural networks Artificial Neural Networks ANNs are mathematical models generally based on concepts learned from the study of animal nervous systems Though they do not accurately model their biological counterparts they still display several of the aptitudes that animal neural networks have evolved over millions of years ANNs are particularly useful in areas such as pattern recognition where traditional statis tical methods of computation have sometimes failed to produce good results Their greatest strength is that given the same general network construction th
74. en reduced to an xyPoint object using methods provided 7 3 The server program The server program is written in C and is able to listen to a specific network port for connection requests from either Subgoal or JSAND applets Once a connection has been established and an appropriate handshake string has been received from the applet the server program attaches the input stream of the socket connection to the UNIX standard output stream The server then calls the main function of the training program and all further output from the training program is forwarded directly to the applet for parsing To achieve this functionality the training program must have its original main function renamed to simulator initialise Additionally since the training program may read configuration information piped from file via the UNIX standard input stream this input stream must be attached to before the simulator initialise callis made The UNIX fork function is used before calling simulator initialise so that the server may 48 7 3 The server program Project Details continue to listen for more applet connections and fork more training programs as necessary The training program is statically linked to the server program at compile time 49 Chapter 8 Critical Appraisal The JSAND IO space display system provides enhanced functionality over previously devel oped IO space display systems JSAND integrates the func
75. er Program The server program must be built along with the training program as described in the maintenence document Assuming that the server is built and ready to run it may be simply started at the Solaris command line by name for example SimServ If the training program requires data to be piped to it the command should also include the file and the pipe command for example SimServ NetStates Once run the server will await network connections and fork a new training program for each connection received from a subgoals applet or a JSAND applet 59 Chapter 11 Maintenance Document The Java parts of the software deliverables have been developed using the Java object oriented philosophy which has allowed most software modules described in the plan to be implemented as separate classes In some cases good object oriented programming practice has led to two modules being contained within the same class For example the neural network data storage module also contains methods to load and save a network functions which were indicated as separate software modules in the plan 11 1 Compilation of the source files The javac compiler may be used to compile both applets and the jsand program Gcc is used to compile the server program with linking options dependent on the training program being linked to Example compilation commands for each program are as follows e The Jsand application may be built by issuing a j
76. ey can be applied to a wide variety of different pattern recognition tasks Whereas statistical analysis techniques used to interpret pattern based inputs would almost certainly require a new implementation for each specific problem Traditional AI approaches to pattern recognition also suffer from this failing In order for an AI rule based search to yield good results for pattern comparison in most cases it would first require hard coding of problem domain specific knowledge before being able to make useful comparisons In any cases the neural network solution will avoid the need for this by being able to learn new patterns to recognise without the need for extensive redesign at the implementation level 2 2 Network Structure Context Survey 2 2 Network Structure Weight values associated with L each of the neurons inputs Inputs from other neurons Neuron processing unit Output Value of neuron Figure 2 1 A neuron processing unit For this project only feed forward strictly layered neural networks will be considered Given an appropriate number of layers a network can be trained to produce a number of desired outputs on recognition of an input pattern Such a network is constructed from neuron units several of which make up a single layer Pattern recognition begins when the input layer of the network is presented with a pattern to process Patterns can be analog or digital with each input being given
77. f supervised learning in an Artificial Neural Network ANN using subgoal chains The software deliverable for this part of the project was to render a two dimensional display of each subgoal to be used in the supervised learn ing session and then be able to superimpose a plot of the actual learning achievement of the network toward each of these subgoals as learning progressed By March 1999 work towards an initial implementation of this subgoal component was well underway and a partially functional demonstration version of the subgoal renderer was submitted to Dr Weir at that time This demonstration software has subsequently been maintained and enhanced by members of Dr Weir s group as an aid in their own research with all changes having been made available to me in February 2000 The second part of this SH project was to develop a two three dimensional display of the 23l 6 1 Specification changes proposed during February 2000 Revised Project Plan Timetable current input output mapping for a given two input ANN Changes have been made to the original planned implementation of this part of the project primarily to allow the entire input output display system to be written in Java The advantages in this approach are described in the following proposed changes section 6 1 Specification changes proposed during February 2000 In the light of changes made to the subgoal component software and discussions of possible new enha
78. for the case when supervised learning takes place in stages each new stage being a subgoal further toward the ideal output goal state 1 1 1 Visualisation of input output space Typically the activation state of each neuron in a multi layer neural network is calculated using a sigmoidal activation function The complete input output I O mapping for such a network can then be found by feeding all possible combinations of input patterns into the network up to a given resolution for real number inputs to find each output pattern exe 1 1 Problem Definition Problem Definition in the mapping This way of finding a networks I O mapping is very time consuming and unsuitable for real time display of a networks current I O mapping Instead this project draws on the theory of hyperbands developed in St Andrews to approximate the sigmoidal activation function in real time This will allow the production of a real time display of the networks I O mapping The use of a semi linear hyperband activation function allows the output responses over and area in input space to be quickly calculated by geometric means This process can be repeated for all areas in the input space and the results can be combined to give an overall I O mapping for the current weight state of that network 1 1 2 Visualisation of subgoal chains The resulting output state of a neural network after training towards a goal output state is not always the best approxima
79. g construction of the Jsand program the step through regions function proved invalu able in graphically debugging the data structures used Being able to display only one region at a time has allowed me to see clearly when incorrect region data was generated The telnet program under solaris was useful in testing the functionality of the server program without using the client applets Telnet allows connections to be made to the server port and textual commands to be issued In this way the correct behaviours of the server on the network could be tested for 061 Chapter 12 Software Listings 12 1 The Jsand Application Jsand java Jis dede dede dede doe eno k k dE dex ak k k k ak k k EE EORR JSand Hyperband display system application build By Duncan McPherson x Sth of March 2000 K ak k ak ak k ak k de ak k ak ak k ak k k k ak K k ak K ak ak k k ak K K ak K k K k K k import java awt event import java awt import java io import java util import java awt image import java lang Math x public class Jsand extends Frame implements ActionListener WindowListener MouseListener MouseMotionListener The display canvas and all widgets needed to control it public IOSpaceCanvas display private Menultem edit load save quit about private Button zoomin zoomout stf stb redraw wintermute island threed private Panel panell private Checkbox drawbands numregions axes stepon bounds Te
80. gain due to the tight specification of the JVM and the feature rich Java OS libraries Also the inherent object oriented structure of Java removing the use of global variables and forcing each software module defined as an object to communicate with other such objects via well defined channels 5 1 3 Integrating Java and C software modules via the JNI Integrating software components implemented in Java to machine specific object code is possible using the JNI Java Native Interface The JNI allows java classes to call native machine specific code and pass data to and from these native modules In the case of this project the main advantage gained through this approach is the use of the XGL rendering libraries available only to native code compiled for Solaris These libraries surpass the primitive graphics functionality of Java in both speed and functionality whilst losing the cross platform advantages inherent in Java To retain maximum portability of the project it is intended that whilst modules concerned with displaying I O space are platform specific to Solaris the remaining Java code will be able to compile and function independently allowing the subgoal chain visualisation functionality to be useable on any JVM 03 5 2 Software modules required Implementation and Testing Plans 5 2 Software modules required 5 2 1 Summary descriptions of core modules XGL Rendering Libraries for Solaris LII 20 8 1 ee I O Spa
81. gb lt 0 rgb 0 if rgb gt 255 rgb 255 g setColor new Color rgb rgb rgb else int on network regionSide region centre Draw the region in red and the outline in black if on 2 0 g setColor Color orange else if on 1 g setColor Color yellow else g setColor Color white g fillPolygon p private void draw_plane double plane Graphics g boolean fill double scale height MAXX MINX zoom Polygon p new Polygon int counter for counter 1 counter lt plane length 1 counter 2 p addPoint width 2 int plane counter scale height 2 int plane counter 1 scale int rgb int plane 0 400 if rgb lt 0 rgb 20 if rgb gt 255 rgb 255 g setColor new Color rgb rgb rgb if fill g fillPolygon p g setColor Color blue if draw bands step g drawPolygon p 101 12 5 The input output display module I0SpaceCanvas java Software Listings private Object translate planes int x int y Object planes double A B C D M N X Y Z yheight int counter vertice point double plane region depthsort new double planes length Object result new Object planes length Find the trigonometric terms used to translate about x and y axes A Math cos Math PI 180 double x B Math sin Math PI 180 double x C Math cos Math PI 180 doubl
82. gions length 1 j display repaint else if source redraw if loader null loader load Thread stop display network new ANN Integer parseInt numhidden getText 10 display bands display network calculateHyperbands display regions display network calculateRegions display translated null display boundaries display network calculateBoundaries display regions display planes display network calculatePlanes display region activation false display repaint 06 12 1 The Jsand Application Jsand java Software Listings else if source threed display planes display network calculatePlanes display threed true display xrot 0 display yrot 90 display translated null display repaint else if source island display regions display network calculateRegions display region activation false display threed false display repaint else if source wintermute display regions display network calculateWintermute display region activation true display threed false display repaint j j public boolean action Event evt Object arg if evt target equals drawbands display draw_bands drawbands getState display repaint return true else if evt target equals numregions display number regions numregions getState di
83. h would have to be interpreted without knowing the position it would be heading toward As this project is experimental further useful display options may come to light during the implementation phase which could improve the users ability to visualise the data being presented Any such additions to the project plan will be detailed in the final project report and a special commentary made of the reasons for an addition 19 Chapter 4 Commentary 4 1 Evolution of the project plan The design and planning of this project to date has been under the supervision of Mike Wier MKW of the Division of Computer Science St Andrews University Week 1 meeting The project title and general overview were proposed by MKW Reference was made to a previous honours project ISLAND Storer 1994 which had a large theoretical overlap with this project Week 2 meeting Initial theoretical groundwork concerning neural networks was discussed with MKW Research references were given along with an overview of the work of several related research efforts in particular ERA Gorse et al 1995 Week 3 meeting Discussion of project theory continued focusing on hyperband theory as developed at St Andrews University A target was set to produce an initial implementation plan for one week hence Week 4 meeting Initial implementation plan was reviewed and a further target was set to produce a document detailing the neural network theory
84. he JSAND applet JsandApplet java Software Listings public class JsandApplet extends Applet implements ActionListener MouseListener MouseMotionListener The display canvas and all widgets needed to control it public IOSpaceCanvas display private Menultem open close quit about private Button zoomin zoomout stf stb redraw wintermute island threed connect disconnect private Panel panell private Checkbox drawbands numregions axes stepon bounds TextField numhidden server Label numlabel boolean laidOut false Communications with applet server variables DatalnputStream is null Socket s null PrintStream os null private ANNReceiver receiver Main entry point to the applet program sets up main display area within the applet viewer No menu bar is created for the applet public void init display new IOSpaceCanvas And an IO space display canvas display network new ANN 1 Create an ANN with x hidden units display bands display network calculateHyperbands display regions display network calculateRegions display boundaries display network calculateBoundaries display regions display draw_bands true this add display West Insets insets insets display reshape insets left insets top 500 500 display addMouseListener this display addMouseMotionListener this Addthe control panel a
85. ight state is unknown a priori and must be searched for from the initial weight state of the network see figure below This search makes use of the known goal output state as a target to train the network towards Travel to a local minimum during this search may produce a poor approximation of the goal weight state so such local minima should be avoided by a good training strategy Initial Weight State Vector of change between initial and subgoal weight states AB AC Weight A Goal Weight State Position unknown until the end of the search Figure 2 6 Changes in weight space during search for a goal weight state Training the network to a number of subgoal output states before the goal output state is an attempt to avoid the problem of local minima preventing the global minimum goal weight state being found A subgoal chain attempts to lead the search to the goal weight state by using nearby subgoals in output space in the hope that each subgoals global minima will also be nearby in weight space The heuristic being used here is that a small change in output space should have a correspondingly small change in weight space It is then hoped 2 5 Non linear subgoal chains Context Survey that a search for a nearby subgoal in weight space is much more likely to succeed than a direct search for a faraway goal weight state 2 5 Non linear subgoal chains A linear subgoal chain sets each subsequent subgoal to b
86. ing packages and also to remote internet browser clients equipped with a suitable VRML plugin In addition to testing the program on the target Solaris platform it should also be possible to have the java only sections of the program functioning on other platforms where only the subgoal chains package is available Further enhancements could allow the java code to be compiled as an applet for use within an internet web page Visualisation of I O space of networks with more than one hidden layer of neurons could be made possible with additional hyperband theory In particular visualisation of networks with two hidden layers would allow a much wider selection of neural network problem I O spaces to be explored The I O space of networks with the topology of 2 n m could also be made possible having a separate I O mapping displayed for each of the m output neurons This would involve having multiple I O space windows displayed simultaneously on screen or a single I O space window intelligently being split into an appropriate number of window panes one for each I O space mapping to be displayed The visualisation of subgoal chains currently only displays the path towards a goal output state within output space It could be possible to also allow the n dimensional path that the network is following within weight space to be plotted using the same theory However since goal weight state is not known the trajectory of such a pat
87. ion function is being used 7 1 4 The load and save functions of the JSAND Application The JSAND program makes use of a file format specified jointly by myself and members of Dr Weir s research group at St Andrews For backward compatibility it is closely related to the file format used by the ISLAND program and so very little conversion work should be needed to allow ISLAND files to be imported into JSAND Descriptions of a neural network are saved as ASCII text with specific tokens being associated with one or more integer and floating point values The textual tokens used are as follows e Neuron i Indicates that a description for a new neuron unit is to follow The integer parameter i is the neuron number 41 7 1 The JSAND input output space display Project Details e Type i Indicates the type of neuron The integer parameter types are 1 for an input neuron 3 for a hidden neuron 2 for an output neuron and 7 for a bias neuron e From Unit i Weight 4f Indicates a weight value joining the current neuron unit to another neuron unit number 96i 96f denotes the real number value of the weight e To Unit i Weight f Indicates a weight value joining two neurons e Done This token is used to indicate the end of a complete neural network description within the file When discovered within a stream the parsar should cause the display to be updated with the new network description Both the parser m
88. ionListener this island new Button Island panell add island island addActionListener this threed new Button Jsand 64 12 1 The Jsand Application Jsand java Software Listings panell add threed threed addActionListener this wintermute new Button Wintermute panell add wintermute wintermute addActionListener this Addthe canvas to the main window and initialise the network numhidden setText 1 display new IOSpaceCanvas And an IO space display canvas display network new ANN 1 Create an ANN with x hidden units display bands display network calculateHyperbands display regions display network calculateRegions display boundaries display network calculateBoundaries display regions display draw_bands true drawbands setState true add display display addMouseListener this display addMouseMotionListener this i public void actionPerformed ActionEvent event Object source event getSource if source quit System exit 0 else if source about 1 AboutDialog a dialog new AboutDialog this About JSAND else if source edit 1 Create an ANN with the number of specified hidden units editor new EditDialog display this else if source load Create an ANN with the number of specified hidden units FileDialog dialog new FileDialog this Loa
89. ionality will be tested individually before the modules are integrated together to form the final package Testing will be done to ensure that error conditions are appropriately handled This will involve giving extreme input data conditions and invalid data conditions to a module and ensuring that it s response is appropriate as detailed in the problem specification Modules will be tested individually with as many valid data sets as development time permits In C based source code macros will be used to assert conditions such as maximum array boundaries and valid parameter ranges Use will be made of preprocessor directives e g the LINE and FILE directives in order to trace faults within modules Java based code will make use of hand coded assertions to help find coding errors Integration of the modules will require C and Java modules to be linked together using utilities provided by the JDK Particular testing focus will be made upon the JNI data communications bridge between the Java and C modules Testing will be done to ensure that data structures and methods are being accessed correctly between modules Finally once successful integration of all modules has taken place the program will be soak tested with as many network configurations as development time permits cT 5 4 Deliverables Implementation and Testing Plans 5 4 Deliverables This project will be formally documented in two parts the plan and the final report I
90. is about menu new Menu About true about new Menultem About new MenuShortcut KeyEvent VK_A about menu add about about addActionListener this menubar add file 63 12 1 The Jsand Application Jsand java Software Listings menubar add about menu setMenuBar menubar Addthe control panel and all its buttons to the main window Panel panell new Panel this add panell West panell setLayout new GridLayout 20 1 zoomin new Button Zoom In panell add zoomin zoomin addActionListener this zoomout new Button Zoom Out zoomout addActionListener this panell add zoomout drawbands new Checkbox Draw Hyperbands panell add drawbands axes new Checkbox Display Axes panell add axes numregions new Checkbox Number Regions panell add numregions bounds new Checkbox Show Boundaries panell add bounds Debugging buttons here stf new Button Step Forward panell add stf stf addActionListener this stb new Button Step Backward panell add stb stb addActionListener this stepon new Checkbox Step Regions panell add stepon numlabel new Label Hidden Units panell add numlabel numhidden new TextField 10 numhidden setEditable true panell add numhidden redraw new Button New Network panell add redraw redraw addAct
91. is environment The main software engineering goals are to provide stable functionality whilst gaining maximum portability of the software to future platforms and maximum performance on the above specified hardware platform 3 2 2 The user The users is assumed to be familiar with the principles of neural networks and also proficient at the use of graphical user interfaces The software tools here are to be used within research situations and as such to be of any use thorough understanding of background theory is required In most cases this software will be used in conjunction with other network training software written by the user for a specific research purpose ae 3 2 Non Functional Requirements Problem Specification 3 2 3 Documentation Full maintenance documentation will be produced for the software for use in later devel opment and alterations as necessary This will describe all software modules produced and their interactions Detailed descriptions will be produced of the algorithms used the core overall structure of each software module and software interfaces between different modules User guide style documentation will be created with a complete and concise description of how to make use of the software via the GUI provided Further documentation will be provided detailing additions to the core specification as described by this plan and also any alterations to the plan that were made in hindsight alo
92. is method information being passed through the network is transformed by the weight values All the networks knowledge is stored in these weights and for the network to learn a new pattern association training must take place to alter the networks weights 2 3 Supervised Learning When the network is untrained its weight values should be set at random Learning of a new input pattern can be accomplished by presenting the input neurons with the new pattern then altering every weight by a general learning rule until the desired output 6 2 3 Supervised Learning Context Survey pattern is produced This is known as a supervised learning strategy where the correct output pattern is provided by the supervisor as a target to alter the weights toward The weight alterations to meet this target are automated during the learning process If a step activation function is employed only a single layer network may be successfully trained using the perceptron learning rule For networks with hidden units further informa tion is needed Finding the amount to change each weight of a hidden layer neuron requires the computation of an error value for hidden units Sigmoidal activation function 0 5 Activating threshold TOARRE Figure 2 5 A sigmoidal activation function A sigmoidal activation function allows such an error called a delta error to be computed from a derivative and used in conjunction with an
93. isplay threed false display repaint else if source island display regions display network calculateRegions display threed false display region activation false display repaint else if source disconnect Disconnect from the server here disconnect 76 12 2 The JSAND applet JsandApplet java Software Listings else if source connect if receiver null 1 receiver receiveThread stop if s null disconnect String host server getText Attempt to connect to the given server name and then involk the ANN simulator try s new Socket host 5555 is new DatalInputStream s getInputStream os new PrintStream s getOutputStream os print ANNSIMSERVER System out printIn Server initialised Wn Create a new thread which handles all output from the newly connected server receiver new ANNReceiver display is numhidden this catch Exception ex System out println Error ex System out println Unable to connect to host n j j j public void disconnect if s null try s close catch Exception ex s null public boolean action Event evt Object arg if evt target equals drawbands display draw_bands drawbands getState display repaint return true else if evt target equals numregio
94. ivated WindowEvent event public void windowOpened WindowEvent event 70 12 1 The Jsand Application Jsand java Software Listings REE PG dope dE k k k k k k doge po p pae k k k ak E k k k k ak e eae ak k k k k ie X Class ANNLoader Purpose Manages all I O between Applet and ANN Simulator and an input file that may contain one or more network descriptions If more than one they are loaded to form an animated sequence K k ak k ak ak k k ak K K k ak k ak K K K ak K k k ak k ak K k k ak K k aK ak K ak EE E RO k K k K k class ANNLoader implements Runnable public Thread loadThread This thread which must be active to recieve bytes IOSpaceCanvas display public String filename public TextField units public ANNLoader IOSpaceCanvas thedisplay String input TextField numhidden Ju Copy the instance of the reciever object and set up buffer variables and flags associated with decoding incoming frames display thedisplay filename input units numhidden load Thread new Thread this load Thread start Entry point for the thread on start of execution public void run try Thread sleep 1000 catch InterruptedException e Try to load a description of a neural network from the stream Vector description null try FileInputStream fin new FileInputStream filename DatalnputStream is new DatalInputStream fin
95. ject Details axes are altered by the distance the mouse has been dragged allowing the user to rotate the display in real time see figure 7 4 The origins of both the 3d IO space display and the 2d screen display are always at the center of the users on screen JSAND window ensuring rotation is always about the visual center of that window The Jsand display mode allows for both wire frame and solid plotting When the 3d regions are plotted solidly the order of plotting becomes important Those regions that are furthest away in Z the axis perpendicular to the users screen must be plotted first so that nearer regions will be plotted on top of them To achieve this type of hidden surface removal each 3d region is given a depth index which is equal to the furthest Z coordinate of any of its vertices All regions are then sorted in order of increasing depth index This provides a solid orthogonal projection of IO space in three dimensions As the Jsand display is orthogonal and not perspective corrected Jsand is able to rotate the display so that it is viewed from above when the Jsand display mode is first selected The overhead view of IO space closely resembles the Wintermute display mode and from there users can rotate the IO space surface at will to better understand its shape As activation within each 3d region behaves linearly the altitude mapped Jsand display accurately reflects the IO space mapping of an ANN when a hyperband activat
96. ld allow viewing of the networks I O space on a remote X terminal The Application Programmer Interface API for XGL is in the form of C libraries which are statically linked at compile time These libraries provide for both 2d and 3d window based graphical displays Detecting user commands to an XGL window is also possible so that interactions with the display can be implemented 5 1 2 The use of the Java JVM and AWT The Java language and associated virtual machine environment holds many advantages for the display a two dimensional subgoal chain representation and also for the editing of a neural network 22 5 1 Software resources to be used during development Implementation and Testing Plans The Java Virtual Machine JVM includes a full window manager the Abstract Window Toolkit AWT which is accessible by the same software Application Programmer Interface API irrelevant to the actual host window manager This allows software written for the AWT to be run on any hardware platform which has an associate JVM In the case of Solaris the AWT would make use of X Windows allowing the same network graphics model functionality as XGL described previously Portability and reusability of Java source code is excellent the former due to Sun Microsys tems standard specification of the JVM which is closely followed by most JVM vendors Reusability of Java source code is also easy relative to other languages such as C or C a
97. lement At 1 weight excitation xy 0 x_weight xy l y weight Hidden h biasWeight if excitation gt kconstant amp amp excitation lt kconstant out excitation 2 kconstant 0 5 Weight Hidden h output Weights element At 0 weight if excitation gt kconstant out Weight Hidden h output Weights elementAt 0 weight Find the activation of the output unit and return this result if out gt kconstant amp amp out kconstant out out 2 kconstant 0 5 else if out lt kconstant out 0 0 else if out gt kconstant out 1 0 return out public int regionSide double xy int h result 0 00 12 3 The neural network storage module ANN java Software Listings double x weight y weight excitation 0 0 for h 0 h Hidden length h x_weight Weight Hidden h inputWeights elementAt 0 weight y weight Weight Hidden h inputWeights element At 1 weight excitation xy 0 x_weight xy l y weight Hidden h biasWeight if excitation gt kconstant amp amp excitation lt kconstant result 2 if excitation gt kconstant amp amp result 2 result 1 return result public Vector splitRegions Vector regions double hbm double hbc int orignum regions size int vertex rgn double x1 yl x2 y2 ix iy grad boolean plabove p2above dou
98. ll Anew subgoal chain display is created using three input nd points a first subgoal weight state a current weight state and the current weight state public SubGoalChain int width SubGoalCanvas thedisplay Initialise the display module to use all of the n dimensions display_width width display thedisplay Given a new point from the simulator adds this to the 115 12 6 The subgoal display applet SubgoalApplet java Software Listings subgoal chain as it stands public void addOutput ndPoint newPoint if start point null start point newPoint else nd outputs addElement Object newPoint display chain outputs display repaint public void addSubGoal double pos int goalNumber if goalNumber 0 if nd_subgoals size lt goalNumber nd subgoals addElement Object new ndPoint pos else ndPoint nd subgoals elementAt goalNumber position pos j display goals goals display repaint public void setGoal ndPoint theGoal 1 System out println Setting the goal goal point theGoal public void update display chain outputs display goals goals display repaint Updates the display with new dimensions and width variables public xyPoint outputs 1 double magnitude angle display_scale display_angle xyPoint result null if nd outputs size gt 0 amp
99. m 10 if neuron max max neuron else if temp startsWith NTYP netDescription addElement Object new String temp temp temp substring NTYP length temp length int ntype Integer parseInt temp trim 10 if ntype INPUTNEURON inputParsed if ntype HIDDENNEURON num hidden if ntype OUTPUTNEURON outputParsed if ntype BIASNEURON biasNeuron neuron else if temp startsWith FROM netDescription addElement Object new String temp else if temp startsWith TO netDescription addElement Object new String temp else if temp startsWith DONE foundDone true 82 12 3 The neural network storage module ANN java Software Listings try temp is readLine catch Exception e temp null If what we have parsed forms a complete network then generate all network data structures that are needed if inputParsed 2 outputParsed 1 num_hidden lt 0 netDescription null return netDescription Constructs a 2 n 1 network data structure from a textual description Makes use of the class wide variables max biasNeuron and hiddenParsed public boolean generateNetwork Vector description Inputs new Neuron 2 Hidden new Neuron num hidden Output new Neuron System out printlIn Maxneuron max Neuron newNeurons new
100. m display zoom 0 25 display repaint else if source stf if display step 1 display stnum 0 display step true stepon setState true else display stnum if display stnum gt display regions length 75 12 2 The JSAND applet JsandApplet java Software Listings display stnum 0 display repaint else if source stb 1 if display step 1 display stnum 0 display step true stepon setState true else display stnum if displaystnum 0 display stnum display regions length 1 j display repaint else if source redraw Create an ANN with the number of specified hidden units display network new ANN Integer parseInt numhidden getText 10 display bands display network calculateHyperbands display regions display network calculateRegions display boundaries display network calculateBoundaries display regions display region_activation false display planes display network calculatePlanes display translated null display repaint else if source threed display planes display network calculatePlanes display threed true display xrot 0 display yrot 90 display translated null display repaint else if source wintermute display regions display network calculateWintermute display region activation true d
101. mute panell add wintermute wintermute addActionListener this Panel panel2 new Panel this add panel2 South connect new Button Connect to Server panel2 add connect 74 12 2 The JSAND applet JsandApplet java Software Listings connect addActionListener this server new TextField 40 server setEditable true panel2 add server disconnect new Button Disconnect panel2 add disconnect disconnect addActionListener this Addthe canvas to the main window and initialise the network numhidden setText 1 drawbands setState true validate repaint public void actionPerformed ActionEvent event Object source event getSource if source quit System exit 0 else if source about AboutDialog a_dialog new AboutDialog this About HyperBands else if source open display network new ANN Integer parseInt numhidden getText 10 display bands display network calculateHyperbands display regions display network calculateRegions display boundaries display network calculateBoundaries display regions display region activation false display repaint else if source zoomout amp amp display zoom gt 0 25 display zoom display zoom 0 25 display repaint else if source zoomin amp amp display zoom lt 25 display zoo
102. n addition these documents the software developed will be demonstrated on a date specified by St Andrews University Computer Science Division Project Plan Due for completion on Friday the 13th of November 1998 at 4 00 pm e Contents e Problem Definition e Context survey e Problem specification e Commentary e Implementation and Testing plans e Project Monitoring Sheet e References Project Report Due for completion on Friday the 16th of April 1999 at 4 00 pm e Abstract e Contents e Project Plan As described above e Project details e Critical appraisal e Testing summary e User manual 28 5 4 Deliverables Implementation and Testing Plans e Maintenance document e Software listings e Status report 29 5 4 Deliverables Implementation and Testing Plans 30 Chapter 6 Revised Project Plan Timetable At the end of meeting 1 on the 4th February 2000 between myself and Dr Mike Weir a set of proposed changes for this Senior Honours SH project were under consideration in the light of new research work undertaken by Dr Weir s group between March 1999 and February 2000 As initially conceived this SH project would involve the development of two distinct func tional components A development plan covering both components was accepted by the Department of Computer Science on November the 13th 1998 The first of the two compo nents was intended to allow the visualization o
103. ncements to the software as it stands the following development work is proposed e Enhancement of the current subgoal display system to allow the use of selected pairs of neural outputs This is in contrast to the demonstration version of March 1999 which would only generate displays from a function of all the current neural output states for subgoals and also the current state of the network e Better integration of the display system with the ANN network simulator program provided by Dr Weir s group so that the display is generated in real time from the output weight states provided by the simulator The March 1999 version of the display system requires recompilation in order to display a different set of weight states e Altering the subgoal chain display system in any way that is necessary to allow seam less integration with the zip model of subgoal learning that has been developed by Dr Weir s group over the past 11 months e Evaluate the resulting system for the potential to be re implemented in the form of a web applet program This would allow subgoal chain displays to be made available to users as part of a web page in a Java enabled web browser This aspect of the project would depend on portability of the network simulator written in C which would have to be able to supply the display applet weight values Initial thoughts are that a client server situation would be suitable with the simulator running on a server machine
104. nd all its buttons to the main window Panel panell new Panel this add panell East panell setLayout new GridLayout 17 1 zoomin new Button Zoom In panell add zoomin zoomin addActionListener this zoomout new Button Zoom Out 73 12 2 The JSAND applet JsandApplet java Software Listings zoomout addActionListener this panell add zoomout drawbands new Checkbox Draw Hyperbands panell add drawbands axes new Checkbox Display Axes panell add axes numregions new Checkbox Number Regions panell add numregions bounds new Checkbox Show Boundaries panell add bounds Debugging buttons here stf new Button Step Forward panell add stf stf addActionListener this stb new Button Step Backward panell add stb stb addActionListener this stepon new Checkbox Step Regions panell add stepon Display type buttons here numlabel new Label Hidden Units panell add numlabel numhidden new TextField 10 numhidden setEditable true panell add numhidden redraw new Button New Network panell add redraw redraw addActionListener this island new Button Island panell add island island addActionListener this threed new Button Jsand panell add threed threed addActionListener this wintermute new Button Winter
105. ng development a different data set has been used The correct operation of the object based neural network data storage module was verified using many debugging assertions and print statements to standard output When a Java print statement encounters an object reference that object s memory reference is printed in place This has made it easy to verify that the object links in the neural network data structures are correct 52 9 1 The parser Testing Summary 9 1 The parser In order to test the functionality of the parser module a variety of test data sets were generated and the parsed output from these sets verified by eye Since the parser could be faced by a number of unexpected input conditions thorough testing of this module was important both at the level of individual functions and as a whole The same core parser functions have been used in the JSAND and SubgoalsApplet programs This code re use should help reduce the number of problems as testing time was able to be concentrated to on the initial implementation of the parser instead of three separate parser programs In the case of JSAND all parser code is held centrally in the ANN java module a source file common to both the applet and application builds 53 Chapter 10 User Manual 10 1 The JSAND Application fe SAND Java Simple Artificial Neural Display Draw Hyperbands Display Axes Number Regions Show Boundaries Step Region
106. ng configuration data from a file allowing the server to operate unattended forking off new training programs to service applet requests 126 Bibliography Horstmann C S and Cornell G Core Java Volume 1 Fundamentals Prentice Hall PTR amp Sun Microsystems Press 1997 Flanagan D Java In A Nutshell O Reilly amp Associates Inc USA 1997 Muller B Reinhardt J Neural Networks an introduction London Springer Verlag 1990 Storer J Island interactive single layer linear activation network display Technical report The School of Computer Science University of St Andrews The North Haugh St Andrews UK 1994 Franklin S Artificial Minds London MIT Press pp 121 140 1995 Gorse T Shepherd A J and Taylor J G The new era in supervised learning Neural Networks 10 2 345 352 1995 Jackson T and Beale R Neural Computing London Instutute of Physics Publishing 1990 127
107. ng use of ancilliary classes neuron and weight to create the data structure 80 12 3 The neural network storage module ANN java Software Listings class ANN private Neuron Inputs public Neuron Hidden private Neuron Output public double bias 3 public double kconstant 2 5 Constructor function used to create the ANN Topology 2 inputs n hidden and one output hidden neurons public ANN int num hidden 1 Initialise a random number generator to create weight values java util Random generator new java util Random System currentTimeMillis Generate two input units since they are input units their inputs are null Inputs new Neuron 2 Inputs 0 new Neuron null generator Inputs 1 new Neuron null generator Generate n hidden units and connect each to the two input units Hidden new Neuron num hidden int counter for counter 0 counter lt num_hidden counter Hidden counter new Neuron Inputs generator Finally generate a single output unit Output new Neuron Hidden generator Tokens used in the file format public final static String NURN _Neuron public final static String NTYP Type public final static String FROM From Unit public final static String TO To Unit public final static String WGHT Weight public final static String DONE
108. ng with the design reasons for making them 3 2 4 Performance constraints In the case of three dimensional displays the target performance is to be able to rotate these displays in real time i e where a frame rate of less than one frame a second is unacceptable performance This should provide the user enough control over the display to adequately visualise the data being presented 3 2 5 Error handling Error conditions raised during the execution of the program should not cause the program to catastrophically fail i e terminate execution immediately or halt the machine in a busy wait loop Instead the user should be given notification that an error has occurred If possible the user will also be given the option to continue the programs execution or in the case of a recurrent or unrecoverable error the user will be able to terminate the program If the user has chosen to terminate the program on an error condition all open files will be closed and control will be passed back to the host operating system in an orderly fashion 18 3 2 Non Functional Requirements Problem Specification 3 2 6 Possible enhancements Output to the ISO standard three dimensional modelling language VRML may also be possible with altitude maps generated for the visualisation of I O space being converted to VRML indexed face sets within textual VRML output This will allow a snapshot of I O space to be exported to three dimensional modell
109. nimation of changing IO space during supervised learning the server side of this system a small C based server has been written which awaits socket connections from remote JSAND applets On receiving a connection request from an applet the server replies with an acknowledgement and then initialises the neural network training program If initialisation was successful all further textual output from the training program is then forwarded to the applet The C based server runs under Solaris an operating environment allowing the output stream from the training program to be easily attached to a socket connection Once a connection has been established and the JSAND applet has begun receiving textual data from the training program the applet begins parsing this data stream The neural network state information conveyed in the stream is in exactly the same format as JSAND application files with any extra information in the stream being ignored Once a complete new weight state for the current network has been received and a Done token detected in the stream the JSAND applet updates the display An example cartoon animation sequence of a neural network IO space during supervised learning is shown in figure 7 5 cc A 7 1 The JSAND input output space display Project Details Jsand java JsandApplet java 1 Network editor module 1 Applet GUI widgets 2 Application GUI widgets menus and dialogues 3 Save and Load Dialogues 2 Client si
110. ns cO 12 2 The JSAND applet JsandApplet java Software Listings display number regions numregions getState display repaint return true else if evt target equals axes display draw axes axes getState display repaint return true else if evt target equals stepon 1 display step stepon getState display repaint return true else if evt target equals bounds 1 display boundary lines bounds getState display repaint return true return super action evt arg public void mouseDragged MouseEvent e if display null display drag e getX e getY public void mouseReleased MouseEvent e if display null display drag finished e getX e getY public void windowClosed WindowEvent event System exit 0 Stubs required by java for event handlers public void windowDeiconified WindowEvent event public void windowlIconified WindowEvent event public void windowActivated WindowEvent event public void windowDeactivated WindowEvent event public void windowOpened WindowEvent event public void mouseClicked MouseEvent e public void mouseMoved MouseEvent e public void mouseEntered MouseEvent e public void mouseExited MouseEvent e public void mousePressed MouseEvent e 78 12 2 The JSAND applet JsandApplet java Software Listings dx des Pe
111. ntly loaded network to be altered or a new network to be created by the user All network weights will be alterable by an appropriate dialogue box along with associated bias properties for each neuron An intuitive slider bar typed interface should be made use of here allowing changes to be made to values easily using a mouse Values should also be able to be manually entered at the keyboard ANN File loader and saver module This module loads a specified file and given that the file has an appropriate neural network file structure it will be parsed and the new network data passed into the network data storage module to become the current network It should also be able to save the current neural network data for later re loading GUI Management of file import and export The module provides dialogue box functionality allowing a file to be chosen from disc and loaded into memory A save as dialogue box will also be provided to allow a pathname to be chosen for the current network to be saved to 5 2 2 Overview of I O space display modules The I O space display modules are to be implemented in C with communication between other core Java based modules via the JNI Hyperband display calculation module All geometric calculations used to determine the location of the hyperbands will be carried out by this C software module All input network data will be taken from the network data storage module and output will be in the form of
112. nto a grid of smaller square S m 7 1 The JSAND input output space display Project Details areas The activation of the current network is then found for each area in the grid and the areas are then coloured a shade of grey corresponding to their activation a real number between zero and one with black for zero and white for one Generation of an IO space display by the Wintermute method is slow however as each square area must be considered individually and its x and y input coordinates fed into the network to compute the output activation see figure 7 2 7 1 2 The Island display mode In Island mode the mathematical characteristics of a semi linear activation function are made use of to calculate the two dimensional IO space display Hyperbands areas in IO space in which the excitation of a hidden layer neuron varies linearly between the constant K are plotted with one hyperband per hidden layer neuron The whole rectangular display area is then split into regions where the edges of hyperbands intersect Hyperbands are represented as two parallel lines in IO space both of which may be represented by the straight line formula y mx c Splitting the display into regions begins with a single region the four sided rectangular display area itself This region is then split into parts wherever the lines of a hyperband intersect it The splitting process continues until no hyperband lines intersect any of the resulting regions
113. odule and the ANN data storage module of JSAND allow neural networks not restricted to a 2 n 1 topology to be loaded and manipulated However because the JSAND IO space display is 2 n 1 specific the parser will reject neural network weight state description files which do not describe a 2 n 1 network If a network description file contains descriptions for more than one complete ANN separated by Done tokens each separate description will be loaded in sequence to form an animation with an 0 5 second gap between frames In this way the complete saved output from an ANN training program can be loaded into JSAND and the changing IO space display viewed as an animation in either Jsand Wintermute or Island modes All of the usual display options are available and may even be changed dynamically during animation for example the current viewing angle or toggeling the display of boundary lines This flexibility is allowed for because the JSAND file loader module is executed as a separate Java thread to the main JSAND application thread The loader thread causes the display to be updated when each frame of the animation has been successfully loaded 7 1 5 The JSAND applet in client mode When JSAND is used in the applet client server configuration all textual output from the C based training program is forwarded to the JSAND applet via a BSD socket connection At 42 7 1 The JSAND input output space display Project Details Figure 7 5 A
114. ppraisal The subgoal display system relies on new theory produced by Dr Weir at St Andrews and as such has no directly related previous work to be compared against The display applet relies on its ability to acquire data in real time from a training program and as the data stream used is constructed from simple ASCII tokens the applet should be easily interfaced to new training programs developed in the future 51 Chapter 9 Testing Summary At implementation each Java class was tested independently as was each method within the classes The ability of the Jsand program to produce three different types of output display has been central to the validation of the code used to generate each different IO space display type Since each display mode Wintermute Island and Jsand is able to render the same data in different styles each style using different sections of source code the correct operation of one display mode could be visually verified by comparing it to the display produced by another mode Particularly the boundary line generation code has benefited from this cross checking When superimposing the boundary line over a Wintermute mode display in all randomly generated test neural networks the line has been seen to be in the correct place The use of random neural networks as test data sets has enabled the testing of JSAND with a wide range of neural network data as each time the JSAND program has been run duri
115. qe k e class ndPoint The position of this point in nd space public double position public ndPoint double pos 1 position pos 118 12 6 The subgoal display applet SubgoalApplet java Software Listings Given two arrays each representing an n dimensional point in space this routine returns the magnitude between the two public static double magnitude double vectorA double vectorB double result 0 0 int index for index 0 index lt vectorA length index result vectorB index vectorA index x vectorB index vectorA index return Math sqrt result Work out the magnitude of a single vector public static double v_mag double vector double result 0 0 int index for index 0 index lt vector length index result vector index vector index return Math sqrt result Given two arrays each representing an n dimensional point in space this routine returns the angle between the two public static double angle double vectorA double vectorB double result 0 0 int index First work out the vector product of the two points for index 0 index lt vectorA length index 4 result vectorB index vectorA index Thendivide the product by the product of the two vector magnitudes and return the inverse cosine of the result return Math acos
116. rent to a user 2 7 Input Output Space The concept of input output mapping is a useful one during supervised learning as it represents the complete set of output responses the network could produce for all possible input states and a given weight state For a network with only one output and two inputs I O space can be represented in 3 dimensions and the I O mapping can be represented as 11 2 8 Approximating I O space using Hyperbands Context Survey a 3 D graph The number of hidden units in the network does not matter here as only the input and output responses are being considered for the current weight state As the weight state of the network changes so too will the I O space mapping An I O space display can then provide a useful indication of the extent to which the learning process has changed the network v 1st Hidden Output Layer Layer s Layer Input A Input B Output Neuron n neurons in each hidden layer Figure 2 10 Low resolution 10 10 representation of an example network I O space Such a graphical representation of a networks I O mapping can be generated by computing the networks response to every possible input permutation to a given resolution for analogue inputs This can be a slow process depending on the number of neuron units within the network and the resolution of the display being produced 2 8 Approximating I O space using Hyperbands For real time display of I O
117. rers ISLAND system Implementation of communications bridge between subgoal chain display and network simulator Java Native Interface Sockets con nection Completed implementation and rudimentary documentation for all neural network data structures as used for Island Submission of this to mkw on Tuesday 22nd February Week 4 21 25th February Work towards an initial submission in Java of a working version of Storer s ISLAND system Possibly at this stage omitting any theoretical enhancements discussed during week 3 how ever with strong focus on having the display GUI parts of this system operating as a web page applet Tentative submission date Monday 28th February Possible further discussions with mkw on the enhancements of Storers ISLAND system for 3d 2d colour displays Week 5 28 3rd March Continued work on Storers system and submission of a working subgoal display system with communications bridge operational so that live data may be accepted from the network simulator without the need for recompilation Hopefully also the subgoal display system will work as a Java applet Submission date Monday the 6th March Week 6 6 10th March Discussion of the submitted subgoal system and continued refinement of the Java ISLAND 34 6 2 Revised software development plan Revised Project Plan Timetable system Discussion of required parts for the SH project write up with a view to submission of a draft write up comple
118. s Hidden Units Figure 10 1 The main JSAND application window Figure 10 1 shows the main JSAND application window On startup a single hyperband LT 10 1 The JSAND Application User Manual from a 2 n 1 ANN where n the number of hidden units is 1 will be displayed The number of hidden units n may be altered by clicking on the hidden units text box 5 in the figure and entering a new value To generate a new random network click the New Network button 6 One hyperband will be displayed per hidden unit When a suitable network is displayed you have a choice of three possible display modes JSAND initially shows an Island mode display as in figure 10 1 where one hyperband 9 and three regions 8 are shown A boundary line 10 is also present The lower region is coloured orange indicating that it is on the off side of the hyperband the upper region yellow as it is on the on side and the region within the hyperband is white l he display mode may be selected using the Island Jsand and Wintermute buttons 7 The display may be scaled in any of the three modes using the Zoom In and Zoom Out buttons 2 Properties of the current display mode can be configured using the four radio buttons 3 below the zoom controls These buttons allow the optional display of the boundary line numbering of the regions hyperbands and axes In certain display modes some of these options are unavailable for example region numbering cannot
119. splay repaint return true else if evt target equals axes display draw axes axes getState display repaint return true else if evt target equals stepon 1 display step stepon getState display repaint return true else if evt target equals bounds 1 display boundary lines bounds getState display repaint return true return super action evt arg 67 12 1 The Jsand Application Jsand java Software Listings public void windowClosed WindowEvent event System exit 0 public void windowClosing WindowEvent event System exit 0 public void mouseDragged MouseEvent e if display null display drag e getX e getY j public void mouseReleased MouseEvent e if display null display drag finished e getX e getY j Stubs required by java for event handlers public void windowDeiconified WindowEvent event public void windowlIconified WindowEvent event public void windowActivated WindowEvent event public void windowDeactivated WindowEvent event public void windowOpened WindowEvent event public void mouseClicked MouseEvent e public void mouseMoved MouseEvent e public void mouseEntered MouseEvent e public void mouseExited MouseEvent e public void mousePressed MouseEvent e class AboutDialog extends Dialog implements ActionListener WindowListener
120. splay scale double display_width ndPoint magnitude goal point position start point position display_angle ndPoint angle goal point position start point position Coordinate array for the start point all subgoals and and the goal result new xyPoint nd subgoals size 4 1 117 12 6 The subgoal display applet SubgoalApplet java Software Listings result 0 start point xy start point goal point display scale display angle display width result result ength 1 goal point xy start point goal point display scale display_angle display width int counter 1 while counter lt result length 1 result counter4 4 ndPoint nd subgoals elementAt counter 1 xy start point goal point display scale display angle display_width i return result 3 CE EN FE EE e Xe ege poke p pe dee a k k k ak k k k k P x Class xyPoint x Purpose Used to represent an xy position on the output display Generated from an ndPoint object ee k k k k k k ak K kk k k ak K k k ak K k K k k ak K k k ak K ea class xyPoint int x int y RA RA He He k k k k dE k k k k k k eee e k k k k k ak k k k ak k k k k e x Class ndPoint x x Purpose Used to represent an n dimensional point and calcualte a lower dimensional represenation of that point as required for subgoal chain display Cotte ee eee ak K k k ak k ak K k k ak K k aK ak K ak K k k ak K k
121. ss Neuron public Vector outputWeights new Vector public Vector inputWeights new Vector public double biasWeight 0 Initialises the neuron data structures and connects weights between this neuron and all of the neurons providing input to it public Neuron Neuron inputs Random generator Generate a new weight connections between this neuron and its inputs if inputs null int counter for counter 0 counter lt inputs length counter inputWeights addElement new Weight inputs counter this generator biasWeight 1 generator nextDouble 2 public Neuron biasWeight 0 01 94 12 3 The neural network storage module ANN java Software Listings public void addWeight Neuron toNeuron double value Weight toAdd new Weight this toNeuron value System out println Trying to add between this and toNeuron value toNeuron input Weights addElement toAdd Function to generate a hyperband given that this is a hidden layer neuron to be written public hBand findHyperBand double bias double k hBand result new hBand if inputWeights size 2 System out println This is not a two input hidden layer neuron result null else Xis assumed to be the first input and y the second double wlx wly wbias wlx Weight inputWeights elementAt 0 weight wly
122. t if btwn biasNeuron newNeurons neuron biasWeight weight j j I return true Yes I know this function is strange workaround for problems with java maths libraries public double findDouble String in if in startsWith 0 in in Double result new Double in if result isNaN result new Double in return result doubleValue Saves a description of a 2 n 1 neural network to file public void saveNetwork String filename System out println save network filename try FileOutputStream fout new FileOutputStream filename PrintStream pout new PrintStream fout Input layer neurons int ncount wcount for ncount 0 ncount lt Inputs length ncount Neuron n Inputs ncount pout println NURN 4 numberNeuron n pout println NTYP INPUTNEURON for wcount 0 wcount lt n outputWeights size weount Weight w Weight n outputWeights elementAt wcount pout println TO numberNeuron w to WGHT w weight 84 12 3 The neural network storage module ANN java Software Listings Hidden layer neurons for ncount 0 ncount lt Hidden length ncount Neuron n Hidden ncount pout println NURN 4 numberNeuron n pout println NTYP HIDDENNEURON for wcount 0 wcount lt ninputWeights size
123. t substring counter counter 1 equals try result value Float valueOf flist substring comma counter trim floatValue catch Exception e result value double 0 0 value comma counter 1 result value Float valueOf flist substring comma counter trim floatValue return result j J JE o e dE ae EE k pq k k k k k k kk k k ak k k k ak k k k k k ak k k k ak k k k k dk x Class SubGoalCanvas Purpose Creates the subgoal display canvas which is used T 12 6 The subgoal display applet SubgoalApplet java Software Listings to form part of the complete applet GUI interface Ae Hee ak k k ae He Se SE ae K k RE DECR EHE SE RICE ak k ak ee e K K aK ak K ak K K e K K k k K k class SubGoalCanvas extends Canvas Private storage of the subgoal data to be displayed public xyPoint chain null public xyPoint goals null public boolean showgoals true public boolean showattained true public int oldwidth 0 public boolean new point false Image to be used in double buffered redraws Image offlmage public boolean dbuffer true Main redrawing procedure public void paint Graphics window update window public void update Graphics window Thentry to draw the subgoal display with what data we have at present Dimension current_size getSize Graphics g String sl goal String s2
124. te or only core sections during week 7 Week 7 13 17th March Demonstration of the Java ISLAND including all enhancements either 2d or 3d output Also demonstration of the subgoal system as a Java applet or in the event of technical difficulties as a Java application running in concert with the simulator Demonstration will take into account initial feedback from mkw s group members in the light of the submission of Monday 6th of March Easter Vacation Consolidate all documentation produced so far into a final SH project report created in LaTeX Week 8 10 14th April Final revisions to both the software and write up of this SH project Week 9 17 21st April Report due Friday 14th of April with acceptance testing beginning Monday the 17th April 85 Chapter 7 Project Details This project provides two separate display systems each of which allows the current state of a neural network to be visualised dynamically during supervised learning Both display systems share some common code in their applet form as they both rely on the same server system to supply them with training data to render The overview of software modules in the implementation section of the plan indicates this overlap in functionality as modules common to both systems are only shown once However apart from using the same server system to source data both display systems are completely separate and will be described here as such 7 1 The JSAND
125. tion of the goal output realisable In many cases the training process fails to find the optimum set of weights the network should have in order that it may produce an output as close to the goal as possible One solution to this is to simply keep re starting the training until a good solution is found Each training session starts with the network in a new random weight state and over several training attempts a closer approximation to the goal output state may be found A different approach is taken for this project in an attempt to successfully train the network to an optimum output state in a single training session Instead of training the network directly to the goal output state a number of subgoal output states are chosen between the goal and the initial untrained output of the network The aim is to place an optimal weight state close to the current weight state through placing a subgoal output state close to the current output state With appropriately chosen sequences it is hoped that training of a network in this way will be much better at finding and keeping in an optimum output state that moves directly towards the goal output state rather than training directly The objective of this project is to be able to visualise the path that the network s changing output state follows under training and to extend this to visualise training with subgoals in 1 1 Problem Definition Problem Definition particular This path is to be plott
126. tionality provided by the Island and Wintermute programs and expands on both with a third analogue three dimensional display method which allows the user to gain a very intuitive feel for the shape of an IO space mapping In comparing JSAND directly with the Wintermute program however it should be noted that all network activation calculations within JSAND are based on a Hy perband activation function whereas Wintermute allows a Sigmoidal activation function to be used as well The original Wintermute program was designed for off line use however whereas JSAND allows a real time view in Wintermute mode to be generated The original ISLAND program provided greater functionality in its ability to edit neural networks directly without training Storer s Island was able to alter the angle of hyper bands displayed in IO space allowing users to construct networks designed to solve specific problems with no further training However some of this functionality can be inherited by JSAND because of the closeness of the two program s file format Minimal changes would be necessary to allow the import of neural networks that had been edited generated by ISLAND allowing them to be displayed in the superior three dimensional JSAND system ISLAND is limited to providing two dimensional displays of network activation and only registers the activity or lack of activity that each neuron provides for a display coordinate 50 Critical Appraisal Critical A
127. to a local disc In submitting this project report to the University of St Andrews I give permission for it to be made available for use in accordance with the regulations of the University Library I also give permission for the title and abstract to be published and for copies of the report to be made and supplied at a cost to any bona fide library or research worker I retain the copyright in this work Contents Problem Definition 1 ll Problem Definition 2 2 6 0455 zoo BRR YR Re 9 E ER Rm 1 1 1 1 Visualisation of input output space lens 1 1 1 2 Visualisation of subgoal chains 2 Context Survey 4 2 1 Introduction to neural networks ac o oea oneta ognara edia ea 4 22 Network Strutture 22291234 6x4 Rh xe EE a eS 5 2 8 Supervised Learning ce rh 6 2 1 Subgoal chains o s coe soa s octi Rm UR Dea ea RR Ron E ee os G 8 2 5 Non linear subgoal chains as soo saas sdi aalma ee 9 2 6 Visualisation of subgoal chains s 9 2T Vee Output Beak ek ue e Xe mp REOR IO cens DR ep e Ee 11 2 8 Approximating I O space using Hyperbands 12 Problem Specification 15 3 1 Functional Requirements 2 00002 eee eee ee 15 3 1 1 Alteration of network properties lll 15 3 1 2 Visualisation of the networks I O space 4 15 3 1 3 Visualisation of Subgoal chains o os ecx sor o ee ee RG 16 4 2 Non Functional Requirements sos or sk ee oro ex YR
128. to extract output states being generated by the ANN simulator program private static final String SUBGOAL SG_ private static final String GOAL O_g_0 private static final String WEIGHT w_c_0O private static final String OUTPUT O_c_0 private static final String DONE Done public SubGoalReceiver SubgoalsApplet thedisplay DatalnputStream input Copy the instance of the reciever object and set up buffer variables and flags associated with decoding incoming frames display thedisplay inputStream input read Thread new Thread this readThread start Entry point for the thread on start of execution 108 12 6 The subgoal display applet SubgoalApplet java Software Listings public void run try Thread sleep 1000 catch InterruptedException e String temp null double val null int lastGoal 1 boolean goal false boolean finished false int iterations 0 while finished try temp inputStream readLine catch Exception ex System out println Error ex Parse the input stream line by line for known tokens if temp null temp temp trim Anew subgoal is found in the input stream if temp startsWith SUBGOAL temp temp substring SUBGOAL length temp length int sgnum Integer parseInt temp substring 0 temp indexOf trim
129. ublic boolean step false public boolean region activation false public int stnum 0 int width height doublescale Specific to the 3d projection funcitonality angles in degrees public boolean threed false public int xrot 0 public int yrot 90 private boolean drag false private int dragx 0 private int dragy 0 private int dragxrot 0 97 12 5 The input output display module I0SpaceCanvas java Software Listings private int dragyrot 0 Image to be used in double buffered redraws Image offlmage public void paint Graphics window update window public void update Graphics window int counter boolean fill double region_centre Dimension current size getSize width current size width height current size height scale height MAXX MINX zoom If there is 3d data which should be plotted ensure it is translated if threed amp amp translated null amp amp planes null translated translate planes xrot yrot planes Set up a graphics buffer to draw the new image into if offfmage null offlmage createlmage width height if offlmage getWidth this width offlmage getHeight this height offlmage createlmage width height Graphics g offlmage getGraphics Colour the background sky blue for a full repaint g setColor Color cyan
130. umber of hidden units in a neural network to be edited weight values to be randomised and the bias level to be changed The cubic spline interpolation module of the subgoals display applet has also not been implemented as under normal training circumstances enough output state data is generated to plot a smoothly curved path between initial and goal states One enhancement to the Java canvas display code of both the JSAND and Subgoals software was to enable the use of double buffered animation Though this has allowed for flicker free rotation and plotting of new subgoals it has been found to prohibit vectored output of the 125 Status Report Status Report display using the appletviewer s postscript print function Since the double buffer code plots the display first into an image buffer the applet viewer generates postscript output using the entire bitmap image buffer A work around for this would be to disable the sections of code allowing this enhancement Since both neural network training programs supplied by the University of St Andrews require configuration data to be input from the standard input stream at startup the server program must be executed with the appropriate configuration data piped as input This has the disadvantage that the server may only be used once and then must be restarted before another applet may make use of that server s training program This could be overcome by the training programs readi
131. will be between the reference vector and the vector of the new output state position as detailed below In the case of linear subgoals where each subgoal is proportionally nearer the goal output 10 2 7 Input Output Space Context Survey Networks output state after training to subgoal zero Cubic spline path e a Goal output state os i x last subgoal Reference vector ir linear to goal output subgoals Final output state of Point chosen on 2d plane network after training for start of reference vector to last subgoal initial untrained output state Figure 2 9 How the resulting 2d data will be plotted on a plane in the case of linear subgoals state all the subgoals will lie on the reference vector between the networks initial untrained output state and the goal state In the case of non linear subgoals each subgoal output state will be determined by the given subgoal generation algorithm As the network is trained to each subgoal the resulting output state can be calculated on the two dimensional output space plane The collection of 2 d points will represent the networks path through output space as subgoal training is in progress A cubic spline curved path plotted between subsequent output states in two dimensions allows the output state changes being displayed to be smoothly represented Such a curved output space path will also help to make trends in changing output state more immediately appa
132. with a cross onscreen private void mark_subgoal int x int y int size Graphics g g setColor Color yellow g fillOval x int size 2 y int size 2 size size g setColor Color red g drawOval x int size 2 y int size 2 size size Adda text label to a subgoal in the display private void label subgoal String label int x int y 114 12 6 The subgoal display applet SubgoalApplet java Software Listings PR E OR e k k REOR k k k k kk k k k k k k ak k k k k k ak k k k ak k k k k e ee debe sedeo k k dO dod ak K op k ak RR OR K k K k k deese e ak K de ede k k int size Graphics g Font f new Font Helvetica Font BOLD 14 g setFont f FontMetrics fm g getFontMetrics f g drawString label x fm stringWidth label 2 y int size 1 4 Class SubGoalChain Purpose Reads and manages the data input from the ANN simulator and generates a subgoal chain using this input class SubGoalChain All class member variables are initialised to a blank canvas state There is no need for a constructor until a connection with the server has been established Arrays used to hold nd and 2d representations ndPoint start_point null Vector nd_outputs new Vector Vector nd_subgoals new Vector ndPoint goal_point null boolean dimensions null int display_width SubGoalCanvas display nu
133. xtField numhidden Label numlabel EditDialog editor 02 12 1 The Jsand Application Jsand java Software Listings ANNLoader loader null Main entry point to the java program sets up main display window and positions it on the screen public static void main String args 1 int gap 200 int start_x 100 int start y 50 Dimension screen Toolkit getDefault Toolkit getScreenSize Jsand window new Jsand window show window setSize screen width gap screen height gap window setLocation start_x start_y The constructor for the main window public Jsand Initial size and position of the display this setSize 800 600 setTitle J SAND Java Simple Artificial Neural Display addWindowListener this First add a menubar to the main window MenuBar menubar new MenuBar Menu file tools about menu file new Menu File true edit new Menultem Edit Network new MenuShortcut KeyEvent VK_E file add edit load new Menultem Load New Network new MenuShortcut KeyEvent VK_O file add load save new Menultem Save Network new MenuShortcut KeyEvent VK_S file add save file addSeparator quit new Menultem Quit new MenuShortcut KeyEvent VK Q file add quit edit addActionListener this load addActionListener this save addActionListener this quit addActionListener th

Project Report - Duncans McPherson`s Home Page

Contents

Download Pdf Manuals

Related Search

Related Contents