User`s Manual PLTool - Planning & Learning research Group
Contents
1. The goal to solve is at obj m1 agencia2 14 For do it I must unload the merchandise m1 of camion1 node 7 As well to unload the merchandise m1 of camion1 the planner must solve the goals at truck camion1 agencia2 inside truck m1 camion1 Prodigy 4 0 chooses the first goal that tries to solve node 8 In order to be able to fulfill this objective can be selected the operator drive truck from agencial to agencia2 node 10 The planner chooses this operator and it s applied The actual state change node 11 But the execution of this operator doesn t have sense because the merchandise m1 have not been loaded yet in the truck Now tries to fulfill the second objective inside truck m1 camion1 node 12 To do that I can apply the operator load truck with the parameters m and camioni in which it s condition that both mer chandise and truck are in the same place node 14 Therefore to obtain this new objective it must be true at truck camion agenciat For do it the planner would have to bring of return camion1 to agen cial If the planner applies the operator drive truck again the camion would be in the agency agencial again but this was true at the begin ning node 18 It will be necessary to return to the last point of decision and to take a different decision than it had been taken Prodigy wil have to return to the decision point in which the truck moved from agencial to agencia2 and to work on2. field prob The system saves the translate problems in the probs directory If you select the option all together the translate problems will be stored in the same file Press the button Translate Problems to translate the pddl problems 3 4 2 IPSS to PDDL The last two text fields are used to specify the domain and the problem set in IPSS that we want to translate to PDDL The ipss temporal dir variable by default is set to instalation directory ipss tmp You can change its value editing the file init lisp in pss directory If you want to translate a set of IPSS problems to PDDL problems you can follow the next steps a First you must create the path ipss temporal dir domain name domain lisp where domain lisp is the file with the defined domain Press the button Find at the right of the text field labeled as Domain Select the directory that contains the file domain lisp that means domain name 27 In this step you must build the problem set as we dicussed in the section 3 2 2 Save this problem set in the path ipss temporal dir domain name probsets probset lisp where probset lisp is the file with the prob lem set Press the button labeled as Find at the right of the text field called Problems Select the file that contain the problem set defined in the last step a Press the button Translate Problems 28 Referencias Aler Mur 2006 Ricardo Aler Mur 2006 Tutori al de Prodigy4 0 h
3. Window that allows the selection of different options for the visu alization from the tree search Best child info If you select this option will print the best children of each node their length to the best solution from them Only success path If you select this option it will print only the success nodes and the failure nodes that diverge one step from the success paths Only tried if t it will print only the expanded nodes a Tree without rules It shows the constructed tree without the fired rules 11 Problem number It allows to select the problem number that we want to visualize Max depth Is the depth to which it should print the tree If we selected the options Only success path and Only tried it obtains the tree that is shown in figure 4 AS SEARCH TREE WITH RULES WINDOW Plan Tree Y finish 30 S finish 4 0 V at obj ml agencia2 5 0 unload truck 6 0 7 unload truck m1 camioni agencia2 7 0 inside truck m1 camion1 8 0 P load truck 9 0 load truck m1 camion1 agencial 10 0 Y LOAD TRUCK 113 Vv at truck camionl agencia2 12 2 V drive truck 13 2 drive truck camionl agencial agencia2 14 2 V DRIVE TRUCK 152 UNLOAD TRUCK 16 nil Figura 4 Constructed tree search for the logistic problem It s necessary to comment that although Only success path had not been selected the result had been the same because Prodigy 4 0 immediately solved the problem without mistakes
4. a goal that consists of having a object in another different airport of which it is now 3 3 Interface for learning macro operators This module allows the learning of macro operators from plans that have been learned previously T he different options that compose this interface can be seen in figure 10 X PLANNING amp LEARNING TOOL DEAK Fie Plan Learn Tools Help Gg B do i PSS i Learn Rules Learn Macros Translators LEARN MACROS install macro First plan step I Printp Last plan step Learn Show Macros Figura 10 Interface for learning macro operators The idea of the macro operators is to memorize useful sequences of action applicable to the planning Therefore a macro operator is a sequence of actions or operators which can altogether be treated Macro operators can be obtained by means of the use of triangular tables and planning The triangular tables are a form of representation of plans An example of triangular table is shown in figure 11 For example in the blocks world there are two operators UNSTACK and PUT DOWN whose definitions can be seen in the file domain lisp of the 23 O1 Precondi operator O1 tions O2 s preconds postconds of O1 operator O2 not established by O1 OS3 s preconds facts in cell above postconds of O2 not established minus delete O2 by O1 or O2 facts in cell above facts in cell above postconds of mimsddmeOS mimedeee Os OB Figura 11 Triangular
5. if before the preconditions in this operator were 13 and inside truck obj lt truck gt at truck truck 10c now we will put them the other way around and at truck truck lt loc gt inside truck obj lt truck gt If we executed the planner again we will verify that the number of nodes is greater in this case 24 nodes This indicates that more nodes than in the previous case have been explored If we unmarked the option Only success paht it s possible to visualize the complete tree search including the failure nodes that have been generated The failure nodes are the red nodes as you can see in figure 6 SEARCH TREE WITH RULES WINDOW Plan Tree Vv finish 30 V o finish 40 7 at obj ml agencia2 5 0 unload truck 6 0 VY unload truck m1 camionl agencia2 7 0 7 at truck camionl agencia2 8 0 drive truck 90 7 drive truck camionl agencial agencia2 10 0 DRIVE TRUCK 110 Y o inside truck m1 camionl 12 0 Vv doad truck 130 load truck mi camionl agencia2 14 nil load truck m1 camionl agencial 15 0 w at truck carnionl agencial 16 0 UO drive truck 17 0 V drive truck camionl agencia2 agencial 18 0 DRIVE TRUCK 19 nil Y inside truck m1 camionl 20 0 V load truck 210 v load truck mi camioni agencial 22 3 Y LOAD TRUCK 233 DRIVE TRUCK 24 2 UNLOAD TRUCK 25 nil Figura 6 Constructed tree search for the logistic problem In summary in this case Prodigy does
6. module is shown in figure 8 In order to make the learning of rules it is necessary to Provide the name of the directory in whom is specified the domain Provide the set of training problems P a Provide a initial set of rules Optional a Establish the search parameters 3 2 1 Defining the domain In this case it will be necessary to do similarly that in section 3 1 2 If the domain already exists we will have to specify to the tool the directory where it is To do that we will press the Find button at the right of the text field labeled with Domain 3 2 2 Set of training problems A set of training problems is necessary to provide to Hamlet P with which it will obtain the set of control rules C To do that it will be necessary to create a directory called probsets in the same directory where we have stored our domain In order to continue with 17 XX PLANNING amp LEARNING TOOL DEAR IPSS Learn Rules Learn Macros Translators LEARN RULES Domain l Buscar Probset l Buscar Rules file la Buscar PARAMETERS 10 V re init learning learning time bound verbosity stop after each problem hes i cost bound E learning type DYNAMIC m eager mode LAZY TOTAL hal Learn Refine Rules Edit Rules Figura 8 Interface for the learning of rules our case of example the subdirectory probsets will be created in the directory logistics Within the directory called probset
7. select the file where we have defined the training set in our case we will select the file problemas lisp 3 2 3 Defining the control rules It is possible to introduce at the beginning of the algorithm a set of control rules instead of starting off of an empty set of rules To do that we will press on the Find button at the right of the File rules text field and we will select the file that stores the control rules 3 2 4 Parameters It is possible to indicate to the algorithm how it must act by means of the establishment of some parameters This parameters are Learning time bound nil or int value It limits the maximum time in seconds in which it looks for the solution By default this value is set to 10 seconds Verbosity in 0 3 It controls the information to print in the ter minal window 0 nothing is written 1 it s written only the resulting plan 2 write information about the nodes and the plan expand 3 write all previous and the control rules fired By default this value is set to 0 Cost type Cost function to use Re init learning Whether you want to continue learning from past experiences or you want to start learning from scratch Stop after each problem Select this option causes that the learning process stops after each problem Cost bound For specifying a cost bound on the solutions Learning type It can be one of e Dynamic Dynamic learning e Deduction Only learning by ded
8. table of operators operator O3 Prodigy directory blocksworld If exist a plan with a sequence of operators like this UNSTACK A B PUT DOWN A The triangular table corresponding to this sequence is shown in figure 12 on A B clear A arm empty UNSTACK A B holding A acres pur vownun clear B arm eimpty on table A Figura 12 Triangular table of operators The generalized triangular table for this example is shown in figure 13 The resulting macro operator UNSTACK amp PUT DOWN would be Preconditions on z y clear z arm empty a adds clear y on table z dels on z y 24 on x y clear x arm empty UNSTACK x y holding x clear y arm empty on table x clear x Figura 13 Triangular table of operators 3 3 1 Parameters The interface allows to specify the steps of the plan among which the corresponding triangular table it s going to built These steps can be intro duced in the labeled text fields with First plan step and Last plan step By default if nothing in these text fields is introduced the system will build the triangular table with the totality of the operators When the program is executed pressing the button labeled as Learn in the domain directory the file macros lisp is created This file contains the learned macro operators If the option Install macro is selected then the system will cr
9. the other goal inside truck m1 camion1 The planner apply the operator load truck with the parameters m1 and camion1 node 23 and later it applies the operator drive truck node 24 that will take the truck from agencial to agencia2 15 3 2 Interface for the learning of control rules As we commented above this module use the system Hamlet that al lows the learning control rules heuristic for Prodigy 4 0 Veloso et al 1996 Hamlet is integrated in the nonlinear planner namely Prodigy 4 0 The inputs to Prodigy 4 0 are Domain theory D or for short domain that includes the set of op erators specifying the task knowledge and the object hierarchy Problem specified in terms of an initial configuration of the world initial state S and set of goals to be achieved G Control knowledge C described as a set of control rules that guides the decision making process The inputs to Hamlet are A task domain D a A set of training problems P A quality measure Q A learning mode L An optimality parameter O The output is a set of control rules C Hamlet has two main modules the Bounded Explanation module and the Refinement module Figure 7 shows Hamlet s modules and their connection to Prodigy Q I M dulo a A Bounded Explanation ST gt i M dulo 2 de Refinamiento ST Reglas de control C ECC C od Prodigy y gt Figura 7 Hamlet The Bounded E
10. Y Sat University Carlos III of Madrid User s Manual PLTool planninglearning LEGAN S September 2006 ndice 1 Introduction 2 2 Installation of the tool 2 3 Graphic interface PLTool 3 al Interface IPSS er xgedeacgh qe kho o OR Seb ORE hos 4 3 1 1 Planner Selection 240 ex 2 4 es Xo 9o Eee xc eee 5 3 1 2 Specification of domains in Prodigy 4 0 5 3 1 3 Definition of planning problems 6 3 1 4 Control rules 225 6444 20 608445 RUSSE RUE 5 9 3 1 5 Parameters x sec ee bre a om RR Eoo RUE RC OG a 9 3 1 6 Execution of the planner 10 3 1 7 Displaying of the constructed plan 10 3 1 8 Displaying of the tree search 10 3 2 Interface for the learning of control rules 16 3 2 1 Defining the domain 17 3 2 2 Set of training problems 17 3 2 8 Defining the control rules 20 3 2 4 Parameters 122 4 res e dow ee R Ghee WE xx 20 3 2 5 Learning of the control rules 21 3 2 6 Refinement of the control rues 21 3 2 7 Edition of the control rules 5 4 424 oo a 21 3 3 Interface for learning macro operators 23 Slo Parameters se s oceno X EE a Y x COE Wee e 25 3 4 Translation interface ee 25 3 4 1 PDDL to IPSS wu kick ns m xe ESO ee 25 34 2 IPSSSQEDDE 453 xen 465 624 ELS 27 1 Introduction The present
11. ble by default is set to instalation directory ipss tmp You can change its value editing the file init lisp in ipss directory If you want to transtate a PDDL domain to IPSS you must follow the next steps You must build the path ipss temporal dir domain name domain pddl where the file called domain pddl contains the domain that we want to translate Press the button Find at the right of the text field labeled as Domain PDDL Select the file domain pddl or the directory that contains it that means domain name a Press the button Translate Domain 26 f all is right you can see a window with the original pddl domain and the translation domain If you want to translate a set of PDDL problems to IPSS problems you must follow this steps a You must build the path ipss temporal dir domain name domain pddl Press the button Find at the right of the text field labeled as Domain and select the file domain pddl or select the directory that contain it domain name a Save the pddl problems that you want to translate in the ipss temporal dir directory as follows ipss temporal dir domain name probs probaxz pddl You can translate one problem or a problem set e If you want to translate one problem press the button Find at the right of the text field labeled as Problems and select the file that contain the problem that you want to translate e If you want to translate a problem set you must type in this text
12. default this value is set to 10 seconds Verbosity in 0 3 It controls the information to print in the ter minal window 0 means nothing is written 1 means only the resulting plan is written 2 write information about the nodes and the plan ex pand 3 write all previous and the control rules fired By default this value is set to 0 Cost type Cost function to use Cost bound For specifying a cost bound on the solutions Multiple solutions If this option is selected the planner doesn t stop when it finds a solution and it will look for all problem solutions Run without rules The execution will be made without loading the specified control rules Backtrack with cost If you selects this option jointly with the option Multiple solutions the planner will show all the solutions with a smaller or equal cost than the last found solution Stop after each problem This option only has effect if a problem set is specified and it causes that the planner stops after each problem Store fired rules The selection of this option makes that the planner store the fired rules in a structure Later these stored rules could be visualized when the tree search is constructed Store plans The selection of this option has similar results to the previous one When it s selected the plans are stored in a structure 3 1 6 Execution of the planner In order to build the plan with the sequence of operators that al
13. document describes the main functions of the tool PLTool a mixed initiative for the automatic planning of tasks developed by the Group of Planning and Learning of the University Carlos III from Madrid This tool is a planner to be run under CLISP Nowadays already exists a plan ner namely PRODIGY that is used interacting with the terminal window PRODIGY is an architecture that integrates planning with multiple learning mechanisms Learning occurs at the planner s decision points and integration in PRODIGY is achieved via mutually interpretable knowledge structures The developed interface will allow to add a graphical support to this plan ner in order to facilitate the interaction between users and system Besides the tool will offer the possibility of working with HAMLET a rules learning system for Prodigy 4 0 now IPSS Hamlet is a system that learns control knowledge to improve both search efficiency and the quality of the solution generated by the non linear planner Prodigy 4 0 2 Installation of the tool You must follow the next steps 1 Create a directory to install the tool For example mkdir pltool 2 Download the files pltool tgz and install sh in this directory 3 Add execute permission to install sh file 4 chmod x install sh 4 Execute the script install sh 3 Graphic interface PLTool Once installed Prodigy 4 0 and Hamlet is possible to start the graphic interface of the developed tool To do that is nec
14. eat two more files the file domain without macros lisp that corresponds with the file of the original domain and the file domain lisp that corresponds with the original domain add the learned macro operators Finally if the option printp is marked the system will print the con structed triangular table in the terminal window 3 4 Translation interface This last module that composes the application is used to translate be tween PDDL and IPSS In figure 14 there are the components of this new module The interface is divided in two parts First of them is destined to the translation from PDDL to IPSS and second one is destined to the translation from IPSS to PDDL 3 4 1 PDDL to IPSS This part allows the translation of domains and problems from PDDL to IPSS The first text field allows the translation of a domain from PDDL to 25 X PLANNING amp LEARNING TOOL Ee IE i IPSS i Learn Rules Learn Macros Translators TRANSLATORS PDDLto IPSS PDDL domain l Buscar 3 Translate domain Domain es Buscar Problems file prefix CS Ia Buscar V alltogether 23 Translate problems IPSS to PDDL Domain a Buscar Probset l Buscar 3 Translate problems Figura 14 Interface corresponding to the translation between PDDL and IPSS IPPS The following two text fields are used to specify the domain and the problem set that we want to translate from PDDL to IPSS The ipss temporal dir varia
15. ective unload truck by other two subobjec tives that the object mi is within the truck camion and that camion1 is in the appropriate place in this case agencia2 It tries to solve the first subgoal as you can see in node 8 The only form to load the merchandise in the truck is to apply the operator load truck Therefore this one is the operator that is applied in node 10 To apply this operator means that the present state changes node 11 The capital letters denote that Prodigy has applied the operator Now Prodigy 4 0 tries to solve the second subobjective that established that the truck is in the transport agency agencia2 that means at truck camion1 agencia2 node 12 In order to be able to add this fact is possible to apply the operator drive truck whose preconditions are true and Prodigy 4 0 can apply it directly node 15 At this moment the truck already is in the agency agencia2 loaded with m1 Therefore the planner can apply the operator unload truck as you can see in node 16 Now the actual state contains the predicate at obj m1 agencia2 and so the problem is solved As we discussed above Prodigy 4 0 constructed the previous tree search without mistakes in the decisions that it tooks Now we are going to force backtracking changing the order of the preconditions of a operator that we have used in the resolution of the problem In particular we will change the order of the preconditions of the operator unload truck it means
16. essary to make the following steps 1 There is a script called pltool sh Add execute permission to this file chmod x pltool sh 2 In order to execute the graphic interface we type pltool sh By default the program has spanish interface If you want an english interface you must edit the file namely znzt lasp in ipss directory and modify the language variable setting its value to en as follows defvar language en setf language en The graphic interface is composes of four different components or modules that implements the main functions of the tool It s possible to interact with each one of these components through a graphic interface in different ways The first interface allows the elaboration of plans The planner by de fault is IPSS or Prodigy 4 0 The second interface allows the learning of rules through a graphic interaction with Hamlet The third interface is used for learning macros a The fourth and last one is used to translate between PDDL and IPSS Next the options of them are detailed X PLANNING amp LEARNING TOOL DER File Plan Learn Tools Gj ayuda IPSS Learn Rules Learn Macros Translators Planner IPSS Domain a Buscar problem dee Rese Buscar PARAMETERS multiple solutions time bound run without rules Searbanky backtrack with costs cost type P J stop afer each problem J store fred rules cost bound Tj store pans Plan Show Pla
17. ge of state They are presented as rules with the form JF THEN Conditions Actions The conditions establish when the operator can be executed and the actions indi cate how to transform the state once the operator has been applied adding or eliminating facts An example of operator in the domain that we are now dealing with can be seen next OPERATOR LOAD TRUCK params lt obj gt lt truck gt lt loc gt preconds lt obj gt OBJECT lt truck gt TRUCK lt loc gt and LOCATION in truck city p lt truck gt lt loc gt and at obj lt obj gt lt loc gt at truck lt truck gt lt loc gt effects O add inside truck obj lt truck gt del at obj lt obj gt lt loc gt In this case the conditions for the operator to be able to execute are that the object that we want to be transported is loaded in the truck and that the truck is in the same location that the object that is wanted to be transported The application of this operator has two consequences on the facts On the one side it adds the fact that it indicates that the object is in the truck and on the other side it eliminates the fact that it indicated that the object was in a location The specification of the rest of operators for this domain can be seen in domain lisp in the logistics directory In order to load a domain in the tool we must specify the directory with the definition of the domain domain lisp In our case we will press the Find bu
18. in the decisions that it took In the figure 5 we can see the tree constructed by Prodigy 4 0 The nodes with something enclosed in brackets indicate an objective that Prodigy tries to solve the nodes with something between lt gt indicate a possible instan tiated operator that Prodigy wants to use to reach the objective In addition if these last nodes are in capital letters it indicates that Prodigy 4 0 has ap plied the operator and therefore has changed the state The arrows indicates the order in that Prodigy 4 0 advances through the nodes We start from the objective to solve at obj m1 agencia2 node 5 Now Prodigy4 0 would look for the operators that allow it to reach that goal In this case there is only one the operator unload truck which allows us to unload the object m1 in agencia2 Therefore to be able to apply the operator unload truck is necessary that their preconditions are fulfilled 12 N5 at obj m1 agencia2 N lt unload truck mI camion agencia gt e N8 inside truck m1 camionl N12 at truck camion agencia2 Zo NiO lt load truck mi camion agencial gt v N14 cdrive truck camion agencia agencia gt NII lt LOAD TRUCK MI CAMIONI AGENCIAI gt EESSSSEEQLLLLLZ ES NIS lt DRIVE TRUCK CAMIONI AGENCIA AGENCIA2 Figura 5 Constructed plan for the logistic problem and inside truck obj lt truck gt at truck truck lt loc gt Prodigy4 0 replaces the obj
19. ject 17767 1 gt lt airport 17767 2 gt true in state at obj lt object 17767 1 gt lt airport 17767 3 gt true in state at airplane lt airplane 17767 4 gt lt airport 17767 3 gt some candidate goals nil type of object lt airplane 17767 4 gt airplane type of object lt airport 17767 3 gt airport type of object lt airport 17767 2 gt airport type of object lt object 17767 1 gt object V THEN then select operators unload airplane Figura 9 Edition of the control rules for the problem miprob lisp of the logistics domain The system have generated two different control rules The first control rule is decide subgoal test sepeb 17766 and it s an example of control rule that chooses a goal between all the set of pending goals or sub goals When Prodigy is expanding a node the behavior by default is to execute the operator which is applicable at a certain moment The control rules can be used to change this behavior by default In this case this rule says that the operator fly airplane that takes an airplane from a city to another one does not execute if an object in the initial airport must be loaded in the airplane that means not to make the flight if the merchandise have not been loaded in the airplane The second generated control rule select unload airplain test sepeb 17767 shows an example of control rule that determines when the operator unload 22 airplain must be selected when we must reach
20. low us to go from the initial state to the final state we press the button labeled as Plan The planner execution generates some statistics files in the directory namely logistics 3 1 7 Displaying of the constructed plan After a brief moment Prodigy will have finished and if we press the Show Plan button we will be able to display the build plan and some statistics For the problem defined in miprob lisp the solution reached for the plan ner is shown in figure 2 Therefore the planner indicates us that first we will have to load the object m1 in the truck camion in the agency agencial to drive the truck until the agency agencia2 and to unload the object there 3 1 8 Displaying of the tree search The Show tree button allows to visualize the tree search constructed by Prodigy When pressing this button appears a new window in which we can select between different options This window with its options is shown in figure 3 The options that can be selected are 10 AS PLANS WINDOW Plans problem 1 plani LOAD TRUCK M1 CAMION1 AGENCIA1 DRIVE TRUCK CAMION1 AGENCIA1 AGENCIA2 UNLOAD TRUCK M1 CAMION1 AGENCIA2 total time 0 012001 total nodes 15 best cost 3 gt Figura 2 Operators that allow to take the object mi from agencial to agencia2 2X PLAN TREE PARAMS best child info _ only success path problem number max depth QD aceptar X Cancelar only tried Figura 3
21. mai objects m1 OBJECT Paris CITY camioni TRUCK agencial agencia2 LOCATION state and at truck camioni Paris at obj m1 Paris loc at agencial Paris loc at agencia2 Paris at obj m1 agenciai at truck camioni agencia1 part of camion1 Paris goal and at obj m1 agencia2 This problem can be stored in a file namely miprob lisp within the direc tory probs in logistics In order to load this file in the tool we will have to press the Find button that is at the right of the text field labeled as Problem However the tool also offers the possibility of loading a problem set For do it you must create a file with the set definition as it s specified in section 3 2 2 This file will be stored in a directory called probsets located in the same directory where it s defined the domain When we start the planner pressing the Plan button will be made the planning of each of the specified problems If we want to visualize the constructed plans for each problem we will press the Show Plans button In this case it s possible that you can only see a plan corresponding to the last problem of the problem set specified If you want to visualize the totality of the constructed plans for each problem that integrate the set previous to the execution of the planner it will be necessary to select the option Store Plans With this selected option we will execute the planner and when the Show Plans button is pressed
22. ned domain logistics 3 1 1 Planner Selection The interface displays at the beginning a pull down menu in which it s possible to select the planner that later will be used By default the plan ner that appears is IPSS or Prodigy 4 0 Prodigy 4 0 is a nonlinear plan ner that uses heuristic It starts in some objectives that are not solved yet and it selects the suitable operators that allow it to reach those objectives Veloso et al 1995 3 1 2 Specification of domains in Prodigy 4 0 For the description of a Prodigy domain is necessary the specification of types of elements and operators In the problem logistics there are four dif ferent types of elements objects transports locations and cities The trans ports can as well be trucks and airplanes The locations within the city can as well be airports and postal offices In Prodigy 4 0 this hierarchy of types is defined as follows ptype of OBJECT top type ptype of CARRIER top type ptype of TRUCK CARRIER ptype of AIRPLANE CARRIER ptype of LOCATION top type ptype of AIRPORT LOCATION ptype of POST OFFICE LOCATION ptype of CITY top type as you can see in the file domain lisp in the directory logistics The type top type is the top level type and from it hangs those types that do not have any superior level OBJECT CARRIER The operators represent the possible changes that can happen between states That means the execution of an operator entails a chan
23. ns Show Search Tree Figura 1 Interfaz IPSS 3 1 Interface IPSS In this interface it appears all necessary for loading domains and prob lems that previously we have defined In addition the interface allows the selection of multiple options that are provided to the planner The interface is shown in figure 1 Next we will describe briefly the main functions of Prodigy4 0 and how the different options can be used the interface is shown in figure 1 For a more detailed description of the main functions of Prodigy 4 0 you can read the online tutorial of Prodigy 4 0 t Aler Mur 2006 that describes very detailed the main functions of this planner A planner is nothing else than a computer program that allows to obtain plans to solve problems The problem solver applies the operators to the This tutorial has been elaborated by Ricardo Aler Mur teacher of the University Carlos III of Madrid initial state to arrive to final states In order to build the tree search with the planner it s necessary to make four fundamental steps a Select the planner that we will use Provide to the tool the domain s name Provide to the tool the problem s name Select the file that contains the control rules which will be applied Optional a Select the parameters The different options that the planner displays will be studied through an example Concretly a problem will be elaborated and solved with the tool using the defi
24. s will be necessary to edit a file with the problems that are going to happen to comprise our training set In fact it will be necessary to create a file by each problem set that is wanted to be used In each of these files it will be necessary to write the problems that conform the training set as follows setf all problems def problem1 def problem2 def problemn where def problem is the definition of the problem setf current problem At this point in our subdirectory probs will be five defined problems pgh1 pgh2 pgh3 test and the one that has been created like example in this manual miprob Therefore we will create the file problemas lisp in the directory probsets that will contain the set of training problems as follows 18 setf all problems Pghl setf current problem create problem name pgh1 objects 5 pgh2 setf current problem create problem name pgh2 objects pgh3 setf current problem create problem name pgh3 objects test setf current problem create problem name pgh1 objects miprob setf current problem create problem name problemai objects Therefore our training set will be formed by five different problems In 19 order to inform to the tool of that this one will be the set of training problems that is used it will be necessary to press the Find button at the right of the text field labeled like Probset and to
25. the set of training problems are defined we press the button labeled as Learn After the execution of the algorithm several files with statistics will be generated in the domain directory Among these generated files there is a file that contains the learned control rules 3 2 6 Refinement of the control rules The refinement of the control rules is made for to remove the superfluous rules that have been generated If you want to use this function press the button labeled as Refine Rules 3 2 7 Edition of the control rules It is possible to edit the learned control rules pressing the Edit Rules button in the interface For the example case the logistics domain and the 21 problem set that was defined previously has been loaded When editing the learned control rules it appears a window as shown in figure 9 X RULES WINDOW DEE Learned Control Rules decide sub goal test sepeb 17766 applicable op fly airplane lt airplane 17766 1 gt lt airport 17766 2 gt airport 17766 true in state at obj object 17766 4 lt airport 17766 2 gt some candidate goals inside airplane lt object 17766 4 gt lt airplane 17766 1 gt type of object lt object 17766 4 gt object type of object lt airport 17766 3 gt airport type of object lt airport 17766 2 gt airport type of object lt airplane 17766 1 gt airplane Vv THEN then sub goal V select unload airplane test sepeb 17767 current goal at obj lt ob
26. the tool will show the totality of the constructed plans 3 1 4 Control rules The control rules are used to guide the search that is carried out by the planner In these control rules the conditions are described in which it are applicable and in case that it are applicable it selects it rejects or it prefers an individual candidate In order to apply a control rule given a set of candidates that can be nodes objectives operators depending on the decision Prodigy 4 0 first select between all the possible candidates executing the applicable selection rules with the purpose of obtaining a subgroup of candidates If no selection rule is applicable then Prodigy 4 0 include all the candidates Later the applicable reject rules are executed to filter this set by explicit elimination of the remaining candidates and finally the preference rules are used to find the favorite alternative The file with the control rules that we want to be considered by the planner during the search must be load pressing the Find button at the right to the text field labeled as Rules File As we discussed above this is an optional field and we will only have to add the file name that contains the control rules in case that we want to use them during the search 3 1 5 Parameters The parameters that can be selected in the construction of the tree search are Time bound nil or int value It limits the maximum time in seconds in which it looks for solution By
27. tton that appears at the right of the text field labeled as Domain and will select to the directory called logistics 3 1 3 Definition of planning problems The problems are specified like a initial state end state pair and we want to find the succession of operators that we must apply to arrive to the final state from the initial state If you want to define a problem you must follow this steps Define the problem name Next the types of objects are declared a Definition of the initial state and the final state Suppose that we defined a problem in which in the initial situation an object exists that is wanted to transport from an transport agency agencia to another transport agency agencia2 by means of the use of a truck In addition all these elements are in Paris The objects can be represented as follows objects m1 OBJECT Paris CITY camioni TRUCK agencial agencia2 LOCATION In addition we know that each one of these objects is in Paris Therefore the initial state is state and at truck camioni Paris at obj mi Paris loc at agencial Paris loc at agencia2 Paris at obj m1 agenciai at truck camioni agencia1 part of camion Paris As we commented above the objetive is to take the object m1 to the transport agency agencia2 Therefore the final state is goal and at obj m1 agencia2 The complete problem is setf current problem create problem name proble
28. ttp scalab uc3m es docweb ia sup practicas anteriores prodigy html Universidad Carlos III de Madrid Veloso et al 1995 Veloso M Carbonell J P rez A Borrajo D Fink E Blythe J 1995 Integrating Planning and Learning the PRODIGY Architecture Carnegie Mellon University Veloso et al 1996 Veloso M Borrajo D 1996 Lazy Incrementar Learn ing of Control Knowledge for Efficienly Obtaining Quality Plans 29
29. uction e Deduction Induction Deduction followed by fast induction 20 e EBL EBL style behaviour It also computes utility and removes control rules with utility under a given threshold e Active Active learning It will call the random generator at every cycle to obtain a new training problem It also saves them The random generator needs a random generator to exist for a given domain and is called with filter problem p t so it will only those problems that are known to be solvable and useful a Eager mode The system can operate in two learning modes eager and lazy In eager mode it generates a positive example from every decision that leads to a solution In lazy mode it generates a positive example only if the decision leads to one of the globally best solutions and it was not the choice selected by the problem solving default heuris tics In addition depending on how the search is made on the search space this can be total or partial Total search requires the expansion of the whole search space and partial search expands the search space up to the time limit Then we have four modes e Lazy total not default and whole search space very lazy e Eager total default and whole search space lazy e Lazy partial not default and partial search space eager e Lager partial default and partial search space very eager 3 2 5 Learning of the control rules In order to obtain the control rules once that the domain and
30. xplanation module generates control rules form a Prodigy search tree These rules might be too specific or too general The Refinement 16 module solves the problem of being too specific by generalizing rules when analyzing positive examples It also replaces too general rules with more specific ones when it finds situations in which the learned rules lead to wrong decisions Hamlet step by step learns and refines control rules converging to a concise set of correct control rules i e rules that are individually neither too general nor too specific SP and S Po are planning search trees generated by two calls to Prodigy s planning algorithm C is the set of control rules and C is the new set of control rules learned by the Bounded Explanation module For each problem p in the set of training problems P Hamlet calls twice Prodigy 4 0 In the first call Prodigy generates a search tree ST by solving P without any control rules exhausting the search space to identify the iptimal solutions Hamlet generates new positive examples from ST as ST was not pruned by any control rules In the second call Prodigy uses the current set of learned control rules C and produces a search tree S7Tc Hamlet identifies possible negative examples from the comparison of the pruned search tree STe with the complete search tree ST Positive and negative examples are used to refine the learned rules to produce the new set of control rules C The interface of this
Download Pdf Manuals
Related Search