The Use of an Integrated Tool to Support Teaching and Learning in
Contents
1. amp Commit if Restore Verify States Figure 2 Substate Editor Tool that it stands on another block Specifying a set of Herbrand interpretations with the expression clear Block on Block Block is not adequate as we assume that any instantiation of a substate definition forms a valid ground substate We require clear Block A on Block Other Block Other using the normal conventions for the unification of parameters At this stage validity checks can be made to ensure that all dynamic predicates appear in one substate definition If a predicate is dynamic then it must appear in at least one of the substate definitions for at least one of the sorts it refers to On the other hand if it appears in a definition for more than one sort the engineer will be warned that the formulation may contain redundant information as discussed above This should provide the student with an initial view as to the adequacy of her selection of predicates to characterise the changes that the objects in the domain undergo 2 1 Capturing Domain Operators The next stage of the knowledge acquisition method and most difficult task for the student is to specify operators representing domain actions Operators in GIPO are conceptualised as sets of parameterised object transitions LHS RHS where the LHS and RHS are substate classes the sort of the object parameter An object transition can have different modalities in an operator2. Domain Specifications Task Specifications OCL PDDL External Planners Le Figure 4 Architectural Breakdown of GIPO Students are supported by an online three part tutorial which introduces them to the subject matter in a step by step fashion by leading them to develop a simple example domain model Part one of the tutorial introduces the flat model where operators are primitive and separate Part two introduces a hierarchical model of plan operators which amounts to a principled approach to HTN planning Finally part three introduces the OpMaker operator learning method Whereas the tutorials lead the student through the features methodically for learning about specific fea tures GIPO has an online hyperlinked user manual For those students who need to dig deeper for example final year project students GIPO also has a language manual which defines the underlying knowledge representation language 11 4 Comparison with other specification modelling tools used in IT Teaching It is common now for undergraduate students to be expected to understand and use a wide range of GUI tools for constructing programs designs and specifications Here we describe some currently in use in particular within our own undergraduate degree and compare and contrast them with GIPO The B Toolkit As with AI the teaching of rigorous specification and development of software is particularly challenging The B toolkit 1 th
3. normally it is necessary which means the LHS is a precondition of the operator and after the operator is executed the object affected will be in the situation specified by a fully instantiated RAS The GIPO operator editor helps the student create a graph representation of an op erator where the nodes are the LHS and RHS states of the object sorts involved in the operator Each such node contains an editable state definition While the use of the Operator Editor is adequate to define operators students have diffi culty primarily due to the possible co designation of variables across the different nodes presented to the user although the underlining and right click mechanism described in the state editor is used Using this manual operator tool illustrates to the student the difficulty of knowledge formulation particularly to do with actions A semi automated knowledge acquisition tool in GIPO is OpMaker 7 this helps the user to create an operator set simply by providing example solution sequences The exercise illustrates some of the concepts of Learning from Examples in ma chine learning in particular inductive generalisation To help explain OpMaker we use a popular planning application that is supplied with GIPO the Lazy Hikers domain Two people hikers go hiking and driving around regions of the Lake District with objects such as tents cars regions and actions such as putdown load getin getout drive unloa
4. domain model a full operator set can be induced with the tool After OpMaker has produced the set of induced operators another learning opportunity for the student is to compare this set with the handcrafted set supplied with GIPO From this stage in process the student learns 1 the difficulty in acquiring knowledge about actions 2 how using machine learning techniques one can potentially avoid the need to hand craft action knowledge 3 the problems and limitations of learning from examples to do with convergence of generalisations the need for knowledge refinement and the importance of good examples in learning The construction of operators provides a good opportunity to compare the planning model with work in formal specification of software For example many of our students used the B tool to create software specifications The pre and post condition version of a GIPO operator and an operation specified in B have great similarities as they both specify deterministic instantaneous actions in terms of predicate descriptions 2 2 Domain Validation Tools in GIPO Continuing the analogy with formal specification of software once the student has built up an initial model of the world it is natural to want to validate it As in formal specification this splits into two kinds of validation Type 1 checking the model for inconsistencies between component parts Type 2 checking the model s accuracy with respect to wha
5. 3885 J loaded tenta car1 keswick Finding state for prevailing object sue Finding state for necessary abject transition for tent loaded tent1 car1 helvelyn lt lt fails to match gt gt loaded tent1 jcar1 keswick XS Ea A a pre noa amp zoomin amp zoom out ox Figure 3 The GIPO Stepper e the reachability analysis tool each substate that is defined for a sort can be indexed against the operators that use them either as consumers or producers The reachability analysis tool reveals to the student substates that cannot be produced and hence could only ever be used in the initial state of an object and substates that cannot be consumed and hence form a dead end for the development of an object These dead end substates are only useful in the development of some other object or are of the kind specified only in a goal condition The reachability tool can be used in conjunction with OpMaker it can indicate if the deduced operators do not give an adequate coverage This is shown by the existence of defined states that are not referenced by any operators e the manual stepper the student can dynamically check a domain is ade quately specified against a set of problems by using the manual stepper The student uses drag and drop to select operators and pop up menus to instan tiate them effectively a
6. The Use of an Integrated Tool to Support Teaching and Learning in Artificial Intelligence T L McCluskey and R M Simpson University of Huddersfield Queens Gate Huddersfield HD13DH UK email r m simpson hud ac uk t l mccluskey hud ac uk Abstract Teaching of knowledge intensive AI is particularly hard as the process of how knowledge is acquired is difficult to grasp without practical experience Ac quiring and using knowledge about actions events processes is especially difficult because of the temporal nature of the subject matter In this paper we report on a tool called GIPO that has been used for teaching AI students the areas of knowl edge acquisition knowledge engineering automated planning and machine learn ing We give a short walkthrough of some of GIPO s functions indicating some of the learning opportunities offered We then compare GIPO with other interfaces used in the computing curriculum We argue that using a high level integrated tool such as GIPO for supporting teaching and learning improves the students learning experience and helps integrate the theory and practice in a range of AI and related subject areas 1 Introduction The teaching of knowledge intensive AI is difficult both from a theoretical and practical perspective because of the peculiar problems to do with acquiring and crafting knowledge bases The process of how knowledge is acquired is not easy for a student to grasp without practical experienc
7. cts object classes sorts predicates constraints states operator schema and tasks There are famil iar point and click drag and drop functions to help the user build up a new domain or reuse existing components 2 validation checks for consistency across parts of the developing domain model Once operator schema have been developed GIPO features a plan stepper which helps the user build up their own solutions to problems in a kind of mixed initiative mode 3 resident plan generation engines and an API for plugging in to third party AI planners A plan animator visualiser displays a planner s solution to a problem in terms of the objects which are effected by the plan This can be stepped through by the user to see the effects of operators on objects and their properties A key design goal in building the tool s interface has been to allow the creation of a specification in terms of images that describe domain structure at a high level of generality The tool takes care of the detail of the syntax of the underlying speci fication making it impossible to construct a syntactically ill formed specification The process of domain model development on which this is based is detailed in the literature see references 6 5 for more details In the next section we give a short walkthrough of some of GIPO s functions indicating some of the learning opportunities offered We then compare GIPO with other interfaces and tool
8. d putup walk sleepin tent They do one leg of a long circular track each day as they get tired and have to sleep in their tent to recover for the next leg Their equipment is heavy so they have two cars which can be used to carry their tent and themselves to the start end of a leg To use OpMaker the student must first create a partial domain model containing objects sorts predicates and state invariants describing the problem domain The student then constructs via a drag and drop process a solution to a pre defined task for instance the following is a solution to the task of doing one leg of the circular track and being ready for the next leg in the morning putdown tentl fred keswick load fred tentl carl keswick getin sue keswick carl drive sue carl keswick buttermere getout sue buttermere carl unload sue tentl carl buttermere putup tentl sue buttermere getin sue buttermere carl drive sue carl buttermere keswick getout sue keswick carl walk sue fred keswick buttermere sleepintent sue fred tentl buttermere The student is encouraged to think of each action in terms of a sentence describing what happens For example in the last action we think of this as Sue and fred sleep in their tent in Buttermere Each action consists of an action identifier followed by a sequence of objects that the action depends on or changes From the input of a plan such as the example above and a partial
9. d integrate aspects of knowledge acquisition knowledge engineering automated planning and machine learning We show how the tool s features supports teaching and the student s learning experience and helps integrate the theory and practice in arange of AI and related subject areas 1 1 History and Overview of GIPO GIPO the Graphical Interface for Planning with Objects 12 pronounced GeePo is the name of a family of experimental tools environments for building planning domain models providing help for those involved in knowledge acquisition do main modelling task description plan generation and plan execution GIPO was an output of the PLANFORM project 10 and has been demonstrated in several major AI conferences for example at ECP 01 in Toledo Spain at ICAPS 03 con ference in Trento Italy and at EKAW 04 in the UK It was used in the tutorial track at AI 2003 in December 2003 at Cambridge and is being entered to the ICAPS 05 knowledge engineering competition at Monterey USA in June 2005 Two versions of GIPO GIPO 1 and GIPO II are available for downloading from the website and new versions are currently being produced GIPO integrates a range of planning tools to help the user explore the domain encoding and determine the kind of planner that may be suitable to use with the domain In particular it has http scom hud ac uk planform gipo 1 graphical tools and visual aids for the input display of obje
10. e of the process As is the case with programming and design it seems that a tools environment that allows the student to effectively apply the theory in a practical scenario is desirable From our experience a useful tool to help in the teaching of AI within the computing curriculum should e connect a range of theory taught during lectures with the application of the theory during practical classes it should support a wide range of the curricu lum as the student has not the time to learn to use many tools e integrate AI with other subject areas taught at advanced undergraduate level in computing e be as simple as possible to use having a familiar look and feel but also be effective in allowing the student to produce non trivial implementations of Al Traditionally AI has been taught within practical sessions by the introduction of declarative programming languages such as Prolog Lisp and Haskell While this satisfies the first two points it is not easy to lead students to build or integrate advanced AI functions from the basis of a programming language The tutor would implement AI algorithms to expose their workings but knowledge intensive issues such as domain modelling would be harder to illustrate In this paper we report on a tool called GIPO that has been used for teaching AI students knowledge engineering for AI Planning for a number of years We argue that GIPO meets the criteria set out above and helps students understand an
11. f domains as shown in the figure in the Internal Representation box To enable GIPO to be used as a general domain modelling tool we have developed translators between our internal language and the planning domain language PDDL 11 The API enables external planning systems to interface to the tools to provide scope for testing and fielding alternative planning algorithms to those internal to GIPO 3 Using GIPO in Teaching and Learning GIPO has been used in the teaching of intermediate and final year undergraduates in both introductory and advanced AI modules It offers a wide range of learning opportunities in AI through knowledge acquisition knowledge formulation vali dation and maintenance of domain models inductive learning and automated plan generation Although numbers of student groups typically 15 20 are too small to make any statistical claims anecdotally GIPO seems to help students seem to integrate AI knowledge learned in lectures and to reach a deeper level of under standing of dry subject matter on say the acquisition of knowledge 10 GIPO Environment Local Static Global Static Dynamic Validation Validation Tools Validation Tools Tools Steppers Editors ae Canonical S Internal Representation Planner l Translators Operator Plan Viewers e 3d Tnducti n Parsers OpMaker Partial orf Validated WTA
12. igure 2 to construct substate definitions For each object sort predicates are selected from the predefined list will char acterise a single substate The process is repeated until all possible substates for the sort are defined The possible unifications of variables of the same sort or between variables belonging to the same path in the sort tree hierarchy can be re stricted using a visual indication of possibly unifying variables as shown in the figure The student selects from a popup menu how an individual variable is to unify with a target variable and if the decision is that they must be distinct then a not_equals clause is generated This strategy for dealing with the unification of variables is pervasive in the GIPO tool set The example in the figure in from the famous Blocks World where one possible state of a block is that it is clear and jState View Editing Sort black fo K Predicates Edit State State Definiti B Allsorts on_block block block on_block B B 5 block on_table block clear B Same 5 gripper cleartblock Different gripped block gripper Optional busy gripper Ee B lt lt Remove Add gt gt Filter By Eirst Reference Only All Referenced All Predicates amp Refresh Selection State Variable Bindings ne B B1 State Variable ID fl Change a Clear E Add amp Update Delete
13. ing 13 for AI planning b PNs emphasise execution simulation rather than just domain modelling c Places in PNs do not map across naturally to the idea of states as parameterised predicates 5 Conclusions and Future Work In this paper we have illustrated the use of the GIPO tool and shown how it helps students apply theory such as in knowledge representation knowledge acquisition and formulation that they have learned during lectures Its interface and under lying language uses the object metaphor similar to other tools that students use in the computing curriculum Students are able to use it both to gain experience of a wide range of AI topics knowledge acquisition automated planning learn ing from examples and to obtain a deep knowledge of topics in these areas For example a student may learn about algorithms for learning from examples and representations for planning operators but without application the knowledge is somewhat stale Using GIPO the student can use the OpMaker tool to induce plan ning operators thus both sustaining their knowledge of these areas and integrating the two together Additionally we have argued that GIPO helps students see the commonalities between AI with other subject areas helping them to integrate new knowledge with other parts of the curriculum We plan to further widen the scope of the GIPO tool Two new versions are cur rently undergoing alpha testing and will be available online
14. l Report CVC TR 98 003 DCS TR 1165 Yale Center for Computa tional Vision and Control 1998 J Hoffmann A Heuristic for Domain Independent Planning and its Use in an Enforced Hill climbing Algorithm In Proceedings of the 14th Workshop on Planning and Configuration New Results in Planning Scheduling and Design 2000 D Liu and T L McCluskey The OCL Language Manual Version 1 2 Tech nical report Department of Computing and Mathematical Sciences Univer sity of Huddersfield 2000 T L McCluskey and J M Porteous Engineering and Compiling Planning Domain Models to Promote Validity and Efficiency Artificial Intelligence 95 1 65 1997 T L McCluskey N E Richardson and R M Simpson An Interactive Method for Inducing Operator Descriptions In The Sixth International Con ference on Artificial Intelligence Planning Systems 2002 T L McCluskey and R M Simpson Knowledge Formulation for AI Planning Proceedings of 4th International Conference on Knowledge En gineering and Knowledge Management EKAW 2004 Whittlebury Hall Northamptonshire UK 2004 Published by Springer in the LNAI series 2004 P F Patel Schneider P Hayes and I Horrocks OWL web ontology lan guage semantics and abstract syntax W3C recommendation 10 February 2004 http www w3 org 2004 OWL 2004 Planform An open environment for building planners http scom hud ac uk planform 1999 15 11 R M Simpson T L McCluskey D Li
15. lasses While there were similarities with GIPO in the initial stages of domain ontology acquisition Protege lacked the facilities to execute the model in any way This is not the fault of Protege but the fact that OWL ontologies are currently capable of encoding only static knowledge Students seemed to find GIPO more satisfying as they could build view validate and then execute the model This ability to involve the model in some kind of constructive operation ie plan generation helped the student to see what the point of the knowledge acquisition process was Petri Nets The Petri Net PN has been a well known formalism used to spec ify and analyse actions in concurrent systems and other related systems since the 1960 s PNs are a popular graphical notation used in teaching as they are relatively easy to understand yet have a formal basis Additionally there are online tools which can be used effectively in practical sessions Superficially there are many ideas in common between GIPO s view of AI Planning and Petri Nets the idea of a token in PNs is very similar to an object instance PNs have transitions al though PN transitions are actually more like instantiated planning operators that parameterised transitions used as the basis of GIPO s operators The main dif ferences between GIPO and PNs are that a PNs are aimed primarily at require ments modelling and capture for real time systems rather than domain modell
16. nowledge is structured around ob ject descriptions their relationships and the changes they undergo as a result of the application of operators during plan execution in contrast to the traditional literal based approach used in Planning languages such as PDDL 3 The student identifies the kinds of objects that characterise the domain and organises them around distinct collections of objects which we call sorts into a hierarchy Object instances for each sort are identified Each object instance in a sort is assumed to have identical behaviour to any other object in the sort To assist in this element of the conceptualisation GIPO provides a visual tree editor Figure 1 Domain checking at this initial stage involves enforcing the tree structure and requiring that node names for sorts and objects are unique The student describes the sorts by identifying predicates that characterise the prop erties of a typical object of each sort and relationships that hold between objects We provide an editor to define predicates by a process of drag and drop from the sort tree previously defined Sorts are static or dynamic depending on whether or not objects in that sort are affected by actions during plan execution ie change state Next the student specifies the constraints on dynamic objects sorts This is done primarily by characterising each valid state of an object of each dynamic sort Typically each member of a sort may be in only one state at a
17. ny time and that during plan execution the object goes through transitions which change its state A substate of an object called here substate to distinguish it from a state of the world which is formed from a collection of substates is the set of all properties and relations referring to that object that are true For each sort s the student uses GIPO to encode state constraints by specifying all the sensible interpretations of predicates describing s The substate definitions derived from this process specify the possible Herbrand interpretations determined by considering the domain and its constraints These form the basis of much of the static validity checks that can be carried out on the completed domain specification For example it the sort is door and predicates are closed locked and unlocked then the student would use the tool to state that the only possible interpretations that can be true are locked and closed unlocked and open unlocked and closed Any other combinations eg open and locked or locked closed and open are excluded The problem of forming a substate definition of a dynamic sort is more complex when there are relational predicates referring to that sort and possibly to other dynamic sorts as well The collection of all such substates for the object will be such that at any instance in time exactly one such substate description will be true of the object The student uses the editor illustrated in F
18. ough built and aimed at commercial applications has been successfully used in teaching for a number of years in our own curriculum It alleviates the onerous task of constructing and discharging proof obligations The student can understand the need for such rigor and view the effects of such proof tools with respect to uncovering and identifying errors without the need for them to produce hand proofs themselves As mentioned above state based software specification languages have many features in common with classical AI planning domain languages These shared features are used to provide useful analogies to the student in particular e the concept of a state the need to construct operators in terms of pre and post conditions and state transitions the underlying assumptions of default persistence and the closed world e the use of invariants to help in validation and model documentation Of course the objectives of both tools are different one to rigorously develop soft ware the other to develop an application that solves planning problems In general the B tool allows the user to input more precise details of a domain and is more meticulous at uncovering errors On the other hand GIPO s range of dynamic tools the stepper and the use of plan generators give it an extra dimension that both stimulates students and allows more scrutiny of the domain model compo nents One can simulate operator execution using the B
19. s used in the computing curriculum the B tool Protege and Petri Nets 2 GIPO Walkthrough We have found it useful to present the student with two paths through the material e an online tutorial on how to construct domains the student is led through a staged method of domain development e analysis and execution of a ready cooked domain model this way the stu dent can at an early stage see the result of domain building being able to bind the model with a planner of choice and being able to solve planning problems E Sort View Expert 55 S car O carl car2 9 S person 8 O fred L O sue Piten O keswick helvelyn O fairfield honister L derwent LS couple O couplel Sort Name B Add Delete Selected Object Name Add Object A Commit b Verify ff Restore Close Figure 1 The Sort Editor We sketch the main steps that the student is led through when using GIPO s on line tutorials This follows a simple knowledge formulation method useful for building up structured planning domain models 6 The central conception used to raise the level of abstraction in domain capture is that planning essentially in volves changing properties and relationships of the objects that inhabit the domain This appeals to computing students intuition and is consistent with their studies in object oriented programming and design K
20. shortly e GIPOIII this version is for domain models that require a much more expres sive representation language that a traditional propositional form While GIPO III is based on the same object centric view of the world as GIPO I and GIPO II it allows the user to model continuous processes and events as well as actions and allows numeric properties of objects to be specified e Object Life History and Generic Type interface this tool forms another knowledge acquisition input into the tool in the same way as OpMaker It allows the student to enter a diagram recording the transitions of objects and automatically create domain operators It also allows the user to re use pre stored object patterns that represent typical dynamic objects For example Lazy Hiking domain object behaviour can be derived from a combination of generic objects we call mobile bistate and portable see 8 for details 14 References 1 2 3 4 5 sy 6 7 8 9 10 B Core UK Ltd http www b core com John H Gennari Mark A Musen Ray W Fergerson William E Grosso Monica Crubezy Henrik Eriksson Natalya Fridman Noy and Samson W Tu The evolution of Protege an environment for knowledge based systems development Int J Hum Comput Stud 58 2003 M Ghallab A Howe C Knoblock D McDermott A Ram M Veloso D Weld and D Wilkins Pddl the planning domain definition language Technica
21. t is being modelled With tools such as GIPO local consistency checks on name uniqueness and hi erarchy definition are automatic when these components are being built Addi tionally Type 1 validation includes checks on global consistency through various forms of static validation Most effective in GIPO are the checks which verify that operator definitions do not compromise the substate definitions Additionally the student has several opportunities to learn about and carry out dynamic validation of both Type 1 and 2 using GIPO Planning with Objects hiking File View Edit Validation Plan Tools Windows eola Help a 8 7 amp 8 s 53 6 te E a E wo Open Save View Print Sorts Predicates States PTrans Tasks Check Algorithm Run FStepper Animator E Stepper Window File Action View amp amp 4 a Load Save Preview Print graphics Printtext Close Task Editing Drawing Canvas Operators List ociTask 10 v load erson Place Tent Can Initial State untoad erson Place Tent can 9 getincCar Piace Person Spk iuc Pasen aticar2 keswit walked couplt outgue keswi outdred kesw up enti kesw put_down erson Place Tent put_up erson Place Tend drive erson Car Place Place2 walk_together TentPlace2 Pers X TO za 5 enti Goal State KAGIPO Information DE walkeaccoupi XX Object s State Property Window Current State of Object tent 1 ISL MARO INES
22. tool and hence this gives a primitive plan stepper but not plan generation as with GIPO Protege 2000 Protege 2 is a well established knowledge acquisition tool which aids the user in building up domain models in description logic As with the B tool 12 it is not designed originally for use in teaching and learning but it has many fea tures which make it usable by non experts Our final year undergraduate students are exposed to it ina module entitled The Semantic Web Protege is quite a gen eral tool here we used it to build up ontologies written in the OWL language 9 Its interface is similar in some ways to GIPO the usual array of GUI features can be used to build up object hierarchies and input propositions and class constraints Protege has many interesting features for the student e there are online tutorials that slowly build up a student s knowledge of rele vant features e a DL theorem prover such as RACER can be hooked up to check classes for consistency ie the definition allows at least one member Also class hierarchies can be re assembled as more properties of the classes are input This relies on the use of subsumption to check whether one class subsumes another e the OWLViz plugin can be used to visualise the class hierarchy of the devel oping DL theory We found this feature crucial for students as after chang ing a theory and re running subsumption the student can see clearly the new c
23. ttempting to solve their own planning problems us ing the model Each operator is applied in the current state to generate the consequent state The student proceeds in this manner to verify that the do main and operator definitions do support the known plans for given problems within the domain The stepper operates as a manual forward planner with results of each object transition caused by an operator shown graphically see Figure 3 e running planning engines the student can of course execute one of the supplied planners within GIPO on specified tasks Such is the intractable nature of planning problems that his has to be carefully controlled by the tutor and GIPO Depending on the planner task combination chosen the solution may not be found for a good period of time GIPO has its own planning engines but third party planners are easy to integrate we often use the FF 4 planer which was a past winner of the International Planning Competition After a planner has returned a solution the student can step through the solution using GIPO s animator tool This takes the results of a planner and produces a graphical representation of the object transitions using the same layout as the stepper We have outlined the main components of GIPO above more details can be found in the AI Planning literature e g 7 The overall architecture of the GIPO is shown in Figure 4 At the heart of GIPO is an object centred internal representation o
24. u and D E Kitchin Knowledge Representation in Planning A PDDL to OCL Translation In Proceedings of the 12th International Symposium on Methodologies for Intelligent Systems 2000 12 R M Simpson T L McCluskey W Zhao R S Aylett and C Doniat GIPO An Integrated Graphical Tool to support Knowledge Engineering in AI Plan ning In Proceedings of the 6th European Conference on Planning 2001 16
Download Pdf Manuals
Related Search