Learning Machines

7/30/2019 Learning Machines

1/6

Machine Learning and Adaptation of Domain Models to SupportReal Time Planning in Autonomous Systems

A. Introduction and Research Context

Background: Simulating low-level cognitive be-haviour, such as reaction to stimuli or autonomicactivity, has been a major focus of research and de-velopment in the autonomous systems (AS) commu-nity for many years. Automated assessment of sen-sor data, and reactive action selection in the form of condition-action pairs, is well developed in roboticand control application areas. In contrast, a charac-teristic of high-level cognitive behaviour is the abil-ity to reason with knowledge of action and changein order to synthesise plans to achieve desired longterm goals. This area is not so well understood, ormanifested in applications of real time dynamic AS.Utilising such reasoning abilities enables an agent tochoose which action to perform to achieve a desiredtask based on a deliberative process involving knowl-edge of the environment, resources, goals, and avail-able actions. The implementation of such high-levelbehaviour has been considered problematic in the AScommunity in the past, regarding both the real timereasoning and knowledge representation aspects as

intractable [34].Control systems in autonomous vehicles, however,such as in exploration robots or space satellites, haveto be capable of deliberative planning and scheduling(P&S) to autonomously accomplish high-level tasks(e.g. collect a rock sample at position X , take aphotograph of constellation Y ). In fact, scientistsat NASA for over 20 years have been developingsystems with such P&S technology for the controlof autonomous vehicles, and have deployed systemswhich can plan the control of spacecraft, generateactivities for uploading to spacecraft, schedule ob-servation movements for the Hubble Telescope, andcontrol underwater vehicles [4, 20, 6, 18]. Researchinto this kind of deliberative planning is often termedarticial intelligence (AI) P&S. The AI P&S researchcommunity has been successful in overcoming someof the theoretical problems to do with computationalcomplexity of generative planning, and scale-up of proposed solutions, which dogged the community inthe last century. This is evidenced by the deploymentof AI P&S technology in a wide range of applica-tions: at this years annual ICAPS event1 the eldedapplications reported included re ghting, satellitecontrol, emergency landing, aircraft repair schedul-ing, workow generation, narrative generation, andbattery load balancing. The event also hosts com-

petitions leading to the development of optimisedplanning tools which can be embedded in applica-tions software.

1 icaps11.icaps-conference.org

Challenges of Fielding AI Planning: The ba-sic challenges of utilising symbolic reasoning systemssuch as deliberative planners within real time AS arewell known, and were neatly summarised some timeago by Woolridge and Jennings [34]:(a) the transduction problem: that of translating the real world into an accurate, adequate symbolic description, in time for that description to be useful; (b) the representation/reasoning problem: that

of how to symbolically represent information about complex real-world entities and processes, and how to get agents to reason with this information in time for the results to be useful.This work and similar publications in the agent com-munity discouraged approaches using symbolic rea-soning, although hybrid approaches have been ex-plored in for example dynamic environments [31],and multi-agents systems [13].

The reasoning problem alluded to in (b) is whatmany in the AI P&S community are aiming to solve,and a measure of their success is the growing rangeof applications alluded to above. It is expected thatthis ongoing research will lead to yet more efficientsolvers, which can accept more expressive input lan-guages. More fundamental, and the subject of thisproposal, is the transduction problem in (a ), whichis connected to the representation issue of (b).

For an articial agent to produce plans and de-cisions rationally, it has to have knowledge of theobjects and the dynamic effects of actions withinits environment. A symbolic representation of suchknowledge is called a domain model , and sepera-tion of the concerns of creating a domain model,and the creation of a planning algorithm, is the ba-sis of what is termed domain independent planning .This is in contrast to specialised or xed goal plan-ning such as path planning, where the separation of knowledge of the domain and planning algorithm isoften blurred. Acquiring, validating and maintain-ing a domain model for the purposes of automatedreasoning is a key research challenge, and has longbeen a limiting factor in the exploitation of domainindepenendemt planning. Currently domain modelsare hand crafted and maintained, whereas in AS theyare required to be automatically learned and subjectto adaptation over run time. The aim of this projectis to draw on recent research advances in AI P&S inworking towards overcoming this research challenge,expressed in the research hypothesis : Automat-

ically learning and maintaining an accurate and adequate domain model for the purposes of high-level reasoning, in particular for the processes of P&S, enables effective, sustained goal-directed be-haviour for real time dynamic AS.

1


2/6

By the end of this project we aim to have demon-strated with a prototype the feasibility of real timedeliberative planning in AS, based on a self-adaptingdomain model. If this challenge is achieved, then itwill open the door to implementing high-level cogni-tive behaviour in real time dynamic AS.

We next survey the state of the art, focusing on

adequacy , that the expressiveness of current domainmodel languages, and accuracy , in particular the useof automated techniques to form and keep up to datethe domain model in the context of verication andvalidation constraints.Domain Model Languages: The control mech-anisms of ASs need to be able to represent and rea-son with rich and detailed knowledge of such phe-nomena as movement and resource consumption inthe context of uncertain and continuously changingenvironmental conditions [12]. Traditionally, phys-ical systems with discrete and continuously-varyingaspects have been represented using the mathemat-ical notion of a hybrid dynamical system. This isa system that has a state made up of a set of realand discrete- valued variables that change over timeaccording to some xed set of constraints. Hybridsystems are used for modelling in applications suchas embedded control systems [5].

The research-led standard domain model languagein planning is PDDL (planning domain descriptionlanguage), which is based around a world view of pa-rameterised actions and states, where it is assumedthat a controller generates a collection of instanti-ated actions to solve some goal posed as state con-ditions. It has been extended to cope with real appli-cations such as crisis management [9] and workow

generation [26], and has versions which can repre-sent time and resources. More expreesive modellinglanguages such as PDDL+ have been developed forapplications where reasoning about processes andevents in a mixed discrete/continuous world is nec-essary [10]. PDDL+ was recently used in an ap-plication for developing multiple battery usage poli-cies [19]. Although PDDL is designed for logical pre-condition achievement, specialist forms of planningcan be incorporated into the language using proce-dural attachment [8]. Using this kind of mechanismlow level planning procedures such as real time pathplanning, which benets from a range of specialisttechniques[21], can incorporated within PDDL.

Despite its widespread acceptability, a seriousproblem with PDDL is that it reects the concerns of those working in generative planning, rather thanthe execution and scheduling orientation of many ap-plications. In contrast, scientists at NASA Amesdeveloped the application-oriented language familiesHSTS [22] and then NDDL [17] for their applicationsin the Space arena. NDDL is fundamentally differentto PDDL in that encodings are based around rep-resentations of objects and object instances, whichpersist in predened timelines of continuous activi-ties. Each activity has a start and end time inter-val (to represent uncertainty of duration), and the

distinction between action and state is effectivelyblurred. Plan generation and execution are thereforelinked to a much greater degree than with PDDL.NDDLs concept of timelines are related to the ideaof crafting abstract plans as in the input languages to

HTN systems [16]. The idea of pre-written hierarchi-cal plans to formulate possible behaviours has longbeen a popular type of formalism in which to encodedynamic knowledge for AI applications. A relatedview of how one could formulate dynamics comesfrom the area of Cognitive Robotics [25], which alsoseeks to emphasise the integration of planning andexecution. The idea here, though, is to start with anaxiomatisation of the application environment usinga variant of situation logic, then hand craft genericplans (so-called action programmimg) from whichconcrete plans can be efficiently derived using de-duction. Systems used in Cognitive Robotics suchas GOLOG require more engineering for individualapplications than in classical planning, but appearmore appropriate for the control of robotics devices.

Another strand of research, closely linked to HTNand practical planning, has focussed on rich plan representations [23, 29, 30]. These representationsare intended for the sharing of plans between agents.The richness of these languages stems from the

underlying ontology that contains generic conceptsfrom the planning domain. They have been used ina number of application domains such as emergencyresponse [24] and personnel recovery [33].

The common role of these rich and expressive lan-guage families is to enable engineers to formulate anadequate representation of structural, dynamic andheuristic knowledge for applications involving actionand change. In real time autonomous systems theselanguages have been used to represent a high levelknowledge layer. The key limitation here is the hand coded nature of this kind of knowledge, and the difficulty of validating the model - all currentapplications rely on teams of knowledge engineersto encode and validate the domain model [15]. Tomeet the challenge of domain modelling in NDDL,recent work by NASA scientists is aimed at devel-oping an interactive domain model editor (IMDE)which uses a simulator to short circuit the loopbetween the model and validation of the model [3].This work also points to the use of machine learningtechniques (some developed by the authors of this proposal) to assist in engineering the model.Another promising method that can be used toautomatically synthesise a planning domain model isto translate from an existing formal model in an ap-plication lanaguage . The ICKEPS-09 competitionwas devoted to this area, with applications includinge-Learning, web services composition, and businessprocessing [27]. While this line of work is importantin the context of embedding planning componentsin applications such as workow planning, this isnot so suitable for AS where no formal model existsa priori. Also, in AS the domain model is subjectto renement and adaptation over time, in orderthat goal directed planning function will remaineffective. We propose to adopt machine learningtechniques to effect both the initial acquisition of the domain model, and its evolution over its lifespan.

Machine Learning of Domain Models: Ma-

chine Learning applied to AI P&S has attracted along history of research, and we point the reader toa recent survey for a full account [14]. There havebeen many events on the subject in recent years in-cluding workshops adjunct to AI international confer-

2


3/6

ences (including ICAPS), and elements of the ICAPScompetition series (ICKEPS/IPC). In the context of domain indepenedent plannng, as well as researchaimed at learning a domain model representing thephysics of the world, much of the machine learningwork is aimed at learning heuristics to make the useof a planning engine more efficient.

Domain model learning can be separated intothree concerns: (i) what language is the learned do-main model going to be expressed in? (ii) what in-puts (training examples, observations, constraints,partial models etc) are there to the learning pro-cess? (iii) what stage is the learning taking place- initial acquisition, or incremental, online adapta-tion? For most work done up to now the answersto (i) are some variant of PDDL forming a domainmodel that can be input to planning engines andto (iii) is initial acquisition. However, adaptationcan be viewed as a special case of initial acqisition,where input to the learning process includes the cur-rent domain model as well as training examples etc,and output is the updated model.

Regarding (ii), systems that learn very expressivedomain models tend to demand most detailed in-put. Work in learning domain models for roboticagents [1, 2] assumes that a training mechanism ex-ists with rich feedback mechanisms. Typically, mucha priori knowledge is assumed, such as predicate de-scriptions of states, and partial or total state in-formation before and after action execution. Withsuch rich inputs, systems such as Amirs SLAF [1]can learn actions within an expressive action schemalanguage.

Some recent work on learning domain model has

concentrated on learning with little or no supplieddomain knowledge. The LAMP system [36] can formsimple PDDL domain theories from example planscripts and associated initial and goal states only. Itinputs object types, predicate specications, and ac-tion headings, and from plan scripts taken from plan-ning solutions, it learns a domain model. The do-main model is synthesised using a constraint solver,inputting two sets of constraints: one set is based onassumed physical, consistency and teleologial con-straints - for example, every action in the exampleplan script adds at least one precondition for a futureaction, actions must have non-empty effects, and soon. The other set of constraints is generated usinga type of associative classication algorithm whichuses each plan script as an itemset, and extracts fre-quent itemsets to make up constraints. While LAMPis aimed squarely at helping knowledge engineers cre-ate a new new domain model, LOCM is an algorithmlearns from plan scripts only [7]. As with ARMS, itoutputs a planning domain theory in a PDDL formatbut it inputs only plan scripts - it does not requirerepresentations of initial and goal states, or any de-scriptions of predicates, object classes, states etc.LOCM has been used in a system that learns to playthe Freecell game by observation, with no a prioriknowledge of the game [7].

There have been several other notable develop-ments in learning in uncertain or partially known do-mains. Reinforcement learning , traditionally usedin single goal or policy learning planners, has recentlybeen developed for symbolic or relational learning,

though its potential for learning full models of thePDDL variety is not yet proven[14]. A promisingapproach towards learning incomplete and uncer-tain domain models is ongoing in the Model-lite project [35]. Here the authors use probabilistic logicas the basis for the language of the learned domainmodel.

B. Summary of Aims and Objectives

To summarise, before AS in real time dynamic appli-cations can attain high-level cognitive skills there arestill major challenges to be overcome in the acqui-sition, validation and adaptation of domain knowl-edge. To be able to perform deliberate reasoning innew or changing domains, we propose that an ASneeds to be able to learn and incrementally adjustits understanding of the world, encapsulated in a do-main model. It needs to ensure the accuracy of theevolving domain model with the help of internal ver-ication checks and external validation constraints.The project aims to work towards the solution of these challenges within a programme involving col-laborator applications (identied as CAs below) putforward by collaborators in the consortium behindthe AIS program call. Hence, we set up the fol-lowing objectives, the achievement of which will bemeasured using the criteria following each one:

1. develop an expressive, hybrid domain model language (here called AIS-DDL) for AS Criteria: AIS-DDL will be a generic languageapplicable to the CAs, containing requiredfeautes such as mixed discrete/continuous do-main knowledge. It will be capable of cap-

turing knowledge about actions and change ata human-understandable level of abstraction,and allow for efficient reasoning as required forlearning, planning and validation.

2. research and develop methods for automated domain model acquisition and online adapta-tion Criteria: the methods will be generic to theCAs, with output consisting of models in AIS-DDL; they will maintain the accuracy and ade-quacy of the domain model, and develop heuris-tic knowledge to support planning functions;

3. determine methods and develop tools for knowledge analysis, verication and valida-tion Criteria: the methods will be able to de-tect inconsistencies in the learned models, de-rive new knowledge, and inform further knowl-edge acquisition and learning cycles. Further,V&V criteria will be in terms humans can under-stand, thus enabling a mixed initiative approachto knowledge engineering where appropriate.

4. deliver a prototype demonstrator system Criteria: The system will exhibit delibera-tive planning within the CAs in a virtual world,and therein demonstrate the efficacy of domainmodel acquisition and online adaptation.

Fit with the AIS programme call : This pro-posal will advance the state of knowledge in fourareas of the Research Interests table:

3


4/6

Model Building and Learning (8.0). The pro-posal concerns the acquisition, learning, valida-tion, maintenance and adaptation of referencemodels (here called domain models)

Planning (4.0). The main role of the domainmodel (referred to in the previous point) is toenable automated planning to achieve desiredgoals.

Structural Awareness and Information Abstrac-tion (3.0). To be able to adapt and changethe domain model requires information inferredfrom sensor data.

Verication and Validation of Autonomous Sys-tems (7.0). The proposed project will con-tribute to this in so far as the V&V of the do-main model and learned knowledge.

This proposed projects research is seen by the pro-posers as fundamental to all the collaborators sce-narios as described in the Call.

C. Method and Technical Plan

This research projects method will be based aroundthe following activities: the creation/acquisition of a simulation environ-ment tailored to each CA (analogous to that pro-posed by Scientists from NASA/JPL [3] to exploremixed initiative knowledge engineering). This willprovide the necessary environment for investigativeresearch for domain model learning and verication

and validation of learned knowledge and synthesisedplans.utilising a hierarchical approach to domain modelconstruction, with abstract symbolic knowledge at ahigh level to enable long term planning, with detailedknowledge at a lower level to enable path planningor manipulator planningthe creation of verication axioms and processesbased on the ontological constraints intrinsic to thedesign of AIS-DDL (analogous to those developedfor PDDL [32])the engineering of a set of immutable validationconstraints capturing the physics of the applicationdomainutilising existing domain model learning tools suchas LOCM and LAMP referred to above, and the richsources of relevant literature, for example the inter-national workshop series on P&S for space2 .

With these developments in place, it will be fea-sible to meet the main challenge of domain modellearning. The learned high level knowledge in AIS-DDL will be translated into the input language of existing planning engines in order to test generatedplans using simulation, and the simulator will be usedas the basis of the subsequent demonstrator system.Project Risks: We identify major risk areas in theproject as (a) feasibility of creating simulations of

CAs (b) poor degree of t between planing technol-ogy and the application requirements (c) difficultyin obtaining and eliciting underlying knowledge.

2 http://www.congrexprojects.com/11c05/

The range of potential CAs (as demonstrated in theprogram call) and the similarity of them to existingplanning developments (eg Mars Rover) mitigateagainst (a). The vast experience of the Proposers inapplying AI P&S, and in the knowledge engineeringaspects in general, will help resolve problems arisingin (b) and (c) by judging what is feasible in terms of

the scope and range of the CAs given the timescale.Finally, the project plan is arranged exibly intosix work packages which are progressive and self contained, meaning that deliverables, which willhave an external impact, are output at each stageof the project.

WP1. Analysis of CAs and State of theArt: Determination and analysis of requirements of the set of CAs which cover the high level planningand decision making function of the AS, drawnfrom members of the AIS consortium. Scope of CAs, and identication of experts, documents andother resources available to be used. For each CA:detemination of required planning function, collationof sample required plans, state representations andsensor information.Distill the state of the art from the literature asapplicable to the case studies. Acquisition andtesting of applicable tools eg specialist and generalplanners, learning tools, with potential for use inthe project.Construction of project web site and considerationof routes to transfer technology and exploit researchoutputs. Consideration of potential for integrationof project results with other funded research in theAIS programme.

Delivered: Agreements on the detail and scopeof the CAs, such as I/O from/to a deliberativeplanning function, and a set of detailed criteriawith which to measure success [D1]; a collection of literature and summary overview of applicable stateof the art in planning and learning techniques[D2],a repository of potentially applicable research tools,project website, and initial report on the integrationof research results within the AIS programme[D3].Evaluation: Scope of CAs to be sufficientlytesting to measure all the planned features of thedomain model language, the learning method,online adaptation, validation etc. The survey willbe of publishable standard, and the tools repository

will be used to demonstrate to collaborators thepotential of current real time planning and learningtechnology.

WP2. Conguration of Simulation Envi-ronment: Using D1, D3 and collaborator resourceswhere applicable, congure or acquire a simulationenvironment, for example based on a virtual worldplatform (such as Second Life), to simulate CAs.Identication of the abstractions made and theeffort required to transfer systems developed in thevirtual world to a real scenario.Delivered: report on abstractions made in the

virtual world[D4]; working application simulator,and well dened interfaces [D5],Evaluation: simulator congured to showcase thechosen CAs, the execution of plans based on learneddomain models, and handling user interaction during

4


5/6

execution; visualisation to satisfy end users

WP3. Planning Domain Model Repre-sentation and Ontology: Utilising D2, gaininsights from the major AI approaches to domainmodel representation (e.g. in classical planning,action programming, constraint-based planning),

and formalisms used in hybrid systems design [5],SAT-based mixed discrete/continuous systems [28],classical-based formalisms [10], and situation-calculus-based work [11]. Clarify the relationshipbetween high level notations and low level reactiveplanning knowledge as used in the CAs, and specifya generic I/O language for the planning component.Combine with insights from D1 and ceate the rstversion of AIS-DDL. Dene a rich ontology of domain independentplanning concepts for representing processes, events,actions, uncertainty, and continuously changingvariables that will provide the abstract vocabularyfor AIS-DDL; Design and implement algorithms that mapsAIS-DDL to known langauges such as variants of PDDL to utilise state of the art planning technology.Delivered: specication of generic planner I/O[D6], AIS-DDL[D7], specication of domain modellanguage ontology[D8], translators[D9].Evaluation: D6 and D7 will t the requirementsof the planning function and model represention(respectively) of the CAs (evaluated by handencodings of collaborator problem domains). D8will be evaluated by peer reviewed publication andin combination with D9 using dynamic testing (inWP4 and WP5).

WP4. Verication and Validation: ThisWP will research and develop methods and toolsfor the verication and validation of AIS-DDLdomain models, resulting in more accurate androbust domain models, and a way of validatingthe doman model learning processes(WP5). Thework will draw on D6,D7 and D8 and relevantliterature [32, 15, 16], and produce tools fora) automated verication analysis: the creationof verication axioms and processes based on theontological constraints intrinsic to the design of AIS-DDLb) automated validation checks: the engineeringand encoding of a set of immutable validationconstraints capturing the physics of each of the CAsc) a visualisation tool to allow users to validate byinspection and manipulate the domain modelsThe outputs of these tools will be used as follows: to provide additional input knowledge during theknowledge acquisition process, and to inform eachcycle of domain model adaptation; to output information relevant for the efficiencywith which planners can solve planing problems,and to provide advice on the best planner to use,and to help optimize the representation to supportefficient automated planning.

to augment learned models with knowledgeuseful for the human user (to make them moreunderstandable and intelligible), and useful forenabling translation to other formalisms; derive new knowledge in terms of domain inde-

pendent features that provide further insights intothe underlying model.Delivered: verication tool [D10], validationtools[D11], visualisation tool [D12], report on spec-ication and computational properties of tools[D13]Evaluation: D10-D12 will be evaluated takinginto account number of errors identied from testscenarios, the quality of the additional knowledgecreated, and the success in integrating the outputwith learning functions in WP5, D13 will be sub-mitted for peer review.

WP5. Machine Learning and Adaptationof Domain Models Utilise D2 and D3 to furtherinvestigate forms of knowledge acquisition andlearning, and methods for domain model creation.Assemble a number of sources of input to machinelearning, for each of CAs: (i) sets of sample infor-mation fused from sensor data (ii) domain invariantinformation (iii) derived information from D10,D11in WP4. Utilise KE tools from D3 as appropriate tocreate sample domain model encodings for the CAs.Utilising D7 (the planning ontology), and insightsfrom the literature e.g. [36, 7]a) create an initial domain model acquisition toolb) based on a), create an adaptation tool forevolving the domain model through its online useDelivered: Hand crafted domain models[D14],learning[D15] and adaptation[D16] tools, report onspecication and computational propoerties of thetools[D17]Evaluation: Learned domain models will becompared to D14; the process of adaptation of domain models will be evaluated operationally

within the demonstrator(WP6), D17 will be sent forpeer review publication.

WP6. Demonstrator Systems, ProjectEvaluation and Exploitation: Developmentof the simulation environment to incorporate au-tonomous behaviour in order to demonstrate systemlearning and adaptation capabilities; extensive test-ing using CAs scenarios; overall evaluation of project;potential for future development including integra-tion with other results in the AIS prorgamme, iden-tication using D3 of effort need to transfer resultsfrom the virtual world to the real, and determinationof exploitation routes of developed technologies.Delivered: nal versions of simulation environ-ments and demonstrator scenarios[D18]; pathway toresearch exploitation document[D19]; nal projectreport[D20]Evaluation: evaluate D18, D19 against successmeasures identied in D1 and take up of researchresults by commercial partners; peer reviewed jour-nal publications derived from D20.

References

[1] E. Amir. Learning partially observable deterministic ac-

tion models. In Proc. IJCAI 05 , pages 14331439, 2005.[2] S. Benson. Learning Action Models for Reactive Au-

tonomous Agents . PhD thesis, Stanford University, PaloAlto, California, 1996.

5


6/6

[3] Bradley J. Clement, Jeremy D. Frank, John M. Chachere,Tristan B. Smith, Keith Swanson. The challenge of grounding planning in simulation in an interactive modeldevelopment environment. In Proc. KEPS Workshop,ICAPS , pages 2330, 2011.

[4] J. L. Bresina, A. K. Jonsson, P. H. Morris, and K. Rajan.Activity planning for the Mars Exploration Rovers. In

Proc. ICAPS , pages 40 49, Monterey, California, USA,2005.

[5] L.P. Carloni, R. Passerore, A. Pinto, and A. Sangiovanni-Vincentelli. Languages and tools for hybrid systems de-sign. 2006.

[6] Steve Chien, Benjamin Smith, Gregg Rabideau, NicolaMuscettola, and Kanna Rajan. Automated planning andscheduling for goal-based autonomous spacecraft. IEEE Intelligent Systems , 13:5055, September 1998.

[7] S.N. Cresswell, T.L. McCluskey, and Margaret M. West.Acquiring planning domain models using LOCM. Knowl-edge Engineering Review (To Appear) , 2011.

[8] P. Eyerich, T. Keller, and B. Nebel. Combining Actionand Motion Planning via Semantic Attachments. In Proc.ICAPS , 2010.

[9] J. Fdez-Olivares, L. Castillo, O. Garcia-Perez, and F. P.Reins. Bringing users and planning technology together:experiences in SIADEX. In Proc. ICAPS , pages 11 20,Cumbria, UK, 2006.

[10] M. Fox and D. Long. Modelling mixed discrete-continuous domains for planning. Journal of Articial Intelligence Research , 27:235 297, 2006.

[11] H. Grosskreutz and G. Lakemeyer. On-line execution of cc-golog plans. In Proc. IJCAI , pages 12 18, 2001.

[12] J. Bresina, R. Dearden, N. Meuleau, S. Ramakrishnan,D. Smith and R. Washington. Planning under ContinuousTime and Resource Uncertainty: A Challenge for AI. InProc. Conference on Uncertainty in Articial Intelligence ,2002.

[13] Rune Jensen and Manuela M. Veloso. Interleaving de-liberative and reactive planning in dynamic multi-agentdomains. In In Proceedings of the AAAI Fall Symposiumon on Integrated , pages 2224. AAAI Press, 1998.

[14] Sergio Jimenez, Tomas De la Rosa, Susana Fernandez,Fernando Fernandez, and Daniel Borrajo. A review of machine learning for automated planning. The Knowl-

edge Engineering Review .[15] D. Long, M. Fox, and R. Howey. Planning domains and

plans: Validation and analysis. In Proceedings of the Ver-ication and Validation in Planning workshop, ICAPS09 ,2009.

[16] T. L. McCluskey, D. Liu, and R. M. Simpson. GIPO II:HTN Planning in a Tool-supported Knowledge Engineer-ing Environment. In Proc. ICAPS , 2003.

[17] C. McGann. How to solve it: Problem solving in Europa2.0. Technical report, NASA Ames, 2006.

[18] C. McGann, F. Py, K. Rajan, J. Ryan, and R. Henthorn.Adaptive control for autonomous underwater vehicles. InProc. AAAI , pages 13191324. AAAI Press, 2008.

[19] M.Fox, D.Long, and D.Magazzeni. Automatic Construc-tion of Efficient Multiple Battery Usage Policies. In Proc.ICAPS, Frieburg, Germany , 2011.

[20] G. E. Miller. Planning and scheduling the hubble spacetelescope: Practical application of advanced techniques.In Articial Intelligence, Robotics, and Automation for Space Symposium , pages 339 343, 1994.

[21] M.Naveed, A.Crampton, D.Kitchin, and T.L.McCluskey.Real-Time Path Planning using a Simulation-basedMarkovian Decision Process. In 31st SGAI International

Conference on AI (to appear) , 2011.[22] N. Muscettola. HSTS: Integrating planning and schedul-

ing. In Intelligent Scheduling , pages 169212. MorganKaufmann, 1994.

[23] A. Pease and T. Carrico. Object model working groupcore plan representation. Technical Report AL/HR-TP-1996-0031, United States Air Force Armstrong Labora-tory, Wright-Patterson AFB, OH, 1996.

[24] S. Potter, A. Tate, and G. Wickler. Using I-X processpanels as intelligent to-do lists for agent coordination inemergency response. In Proc. 3rd Information Systems for Crisis Response and Management (ISCRAM) , 2006.

[25] R. Reiter. Knowledge in Action. Logical Foundations for Specifying and Implementing Dynamical Systems . MITPress, 2001.

[26] A. Riabov and Z. Liu. Scalable planning for distributedstream processing systems. In Proc. ICAPS , Cumbria,UK, 2006.

[27] Roman Bartak, Simone Fratini, and Lee McCluskey. Thethird competition on knowledge engineering for planningand scheduling. AI Magazine, Spring 2010, 2010.

[28] J.A. Shin and E. Davis. Processes and continuous changein a sat-based planner. Articial Intelligence , 166, 2005.

[29] A. Tate. Roots of SPARshared planning and activityrepresentation. The Knowledge Engineering Review , 13,1998.

[30] A. Tate. < I-N-C-A> : A shared model for mixed-initiativesynthesis tasks. In Gheorghe Tecuci, editor, Proc. IJCAI Workshop on Mixed-Initiative Intelligent Systems , pages125130, 2003.

[31] A. Walczak, L. Braubach, A. Pokahr, and W. Lamersdorf.Augmenting bdi agents with deliberative planning tech-niques. In in The Fifth International Workshop on Pro-gramming Multiagent Systems (PROMAS-2006 , 2006.

[32] G. Wickler. Using planning domain features to facili-tate knowledge engineering. In Proc. KEPS Workshop,ICAPS , 2011.

[33] G. Wickler, A. Tate, and J. Hansberger. Supporting col-laborative operations within a coalition personnel recov-ery center. In Proc. 4th Knowledge Systems for CoalitionOperations (KSCO) , pages 1419, 2007.

[34] M. Wooldridge and N. R. Jennings. Intelligent agents:Theory and practice. The Knowledge Engineering Re-view , 10(2):115152, 1995.

[35] S. Yoon and S.Kambhampati. Towards model-lite plan-ning: A proposal for learning & planning with incompletedomain models. In Proc. Workshop on AI Planning and Learning, ICAPS , 2007.

[36] Hankz Hankui Zhuo, Qiang Yang, Derek Hao Hu, and Lei

Li. Learning complex action models with quantiers andlogical implications. Articial Intelligence , 174(18):1540 1569, 2010.

6

Learning Machines

Documents

Transcript of Learning Machines