Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence Model Update for Automated Planning Maria Viviane de Menezes and Leliane Nunes de Barros Department of Computer Science Institute of Mathematics and Statistics University of São Paulo, Brazil mariavm,leliane@ime.usp.br Introduction A limitation of this approach is that CTL temporal logic, which is the main formalism used in planning as model checking, does not take into account the actions behind the state transitions and therefore this logic is not appropriate to deal with most of the planning problems. In order to overcome these limitations, Pereira and Barros (2008) proposed an extension of CTL, called α-CTL, whose semantics considers the actions behind the transitions. They have also defined a planning as model checking approach based on α-CTL that is able to cope with complex goals. Given that α-CTL semantics is an action labeled transition system that can be induced by a planning domain specification, can this formal approach be used to develop useful Knowledge Engineering tools for planning? Given the planning goal ϕ and ϕ , this PhD proposal aims to develop a KE tool based on α-CTL logic that is able to: Model checking is a formal approach that consists of automatically solving the problem of verifying if M satisfies ϕ, where M is a formal model of a system and ϕ is a formal specification of a temporal property (Figure 1). A model checker will return success if M satisfies ϕ, otherwise, it will return a counter-example, e.g., a set of states where ϕ is not satisfied. model M property ϕ model checker success or counter-example Figure 1: Model checking functional diagram. Automated planning (Ghallab, Nau, and Traverso 2004) is the artificial intelligence field that studies the process of reasoning about actions and goals, seeking for implementation of planners. Essentially, a planner is an algorithm that synthesizes a plan of actions by analyzing a formal specification of the environment’s dynamic and the agent’s goal. Since the trajectory of a plan have temporal properties that can be formally specified, formal methods based on model checking techniques have been used to guarantee the reliability of plans (Cimatti et al. 1997; Cimatti, Roveri, and Traverso 1998; Daniele, Traverso, and Vardi 1999). A planning algorithm based on model checking (Figure 2) receives a planning domain M , a planning goal ϕ and either returns a plan, i.e., a substructure M ⊆ M that satisfies the goal ϕ; or failure if there is not such a plan. Traditionally, the planning domain is represented by a Kripke structure (Kripke 1963) and the goal is specified in the Computational Tree Logic - CTL (Clarke and Emerson 1982). domain M goal ϕ planner • modify an incorrect plan in order to satisfy ϕ (plan repair). • modify a plan that satisfies ϕ, to also satisfy ϕ (planning for complex goals). • improve the quality of a suboptimal plan for ϕ.(plan optimization). • modify a planning domain when no plan can be found to satisfy ϕ and return such plan (knowledge engineering for planning). Background and Progress to Date Model update is a technique that extends model checking functions in order to support the repair of a faulty system. Zhang and Ding (2008) proposed a model update algorithm (Figure 3) that takes a Kripke model M, which does not satisfy an arbitrary CTL formula ϕ, and generates an updated model M that: (i) satisfies ϕ and (ii) has a minimal change with respect to the original model. To generate M , this approach uses primitive update operations such as: add a relation element, remove a relation element, change labelling function on one state, add a state and remove a state. plan or failure model M Figure 2: Planning as model checking. property ϕ c 2011, Association for the Advancement of Artificial Copyright Intelligence (www.aaai.org). All rights reserved. CTL model update model M’ Figure 3: CTL Model update (Zhang and Ding 2008). 1841 the size of the induced planning domain (M ). One possible way to deal with this is to work directly with the action description (i.e., applying symbolic techniques). We also intend to perform empirical experiments using domains from the International Planning Competition (IPC6 2008). By analyzing (i) different plans generated by outof-shelf planners and (ii) planning problems for which no planner is able to solve, we expect to be able to propose updates either in the plans or in the planning domains. Acknowledgements. This work has been supported by FAPESP project 2010/10845-0. Based on Zhang and Ding ideas, we propose an α-CTL model update approach to perform knowledge engineering for planning. The proposed system (Figure 4) receives: (i) a set of actions A (used to induce the planning domain MA ); (ii) a partial model M, which is a substructure of the planing domain (M ⊆ MA ) and represents the plan that we want to update; and (iii) an α-CTL formula ϕ representing a planning goal. The system will perform modifications in the input model M to (Figure 4): 1. update M generating a plan M that satisfies ϕ or 2. update the planning domain MA in order to get such a plan, if there is no plan that satisfies ϕ, generating a new planning domain MA and a new set of actions A . model M actions A property ϕ α-CTL model update References Cimatti, A.; Giunchiglia, F.; Giunchiglia, E.; and Traverso, P. 1997. Planning via model checking: A decision procedure for AR. In ECP, 130–142. Cimatti, A.; Roveri, M.; and Traverso, P. 1998. Strong planning in non-deterministic domains via model checking. In AIPS, 36–43. Clarke, E. M., and Emerson, E. A. 1982. Design and synthesis of synchronization skeletons using branching-time temporal logic. In LNCS - Logic of Programs, Workshop, volume 131, 52–71. London: Springer-Verlag. Cresswell, S.; McCluskey, T.; and West, M. 2010. Acquiring planning domain models using LOCM. Knowledge Engineering Review. Daniele, M.; Traverso, P.; and Vardi, M. Y. 1999. Strong cyclic planning revisited. In 5th European conference on planning, 35–48. Ghallab, M.; Nau, D.; and Traverso, P. 2004. Automated planning: Theory and practice. USA: Morgan Kaufmann Publishers Inc. IPC-6. 2008. The 6th International Planning Competition. http://ippc-2008.loria.fr/wiki/ index.php/Main_Page. Kripke, S. 1963. Semantical considerations on modal logic. J. Acta Philosophica Fennica 16. Menezes, M. V.; Pereira, S. L.; and Barros, L. N. 2010. Model Updating in Action. In Workshop on Knowledge Engineering for Planning and Scheduling – ICAPS. Menezes, M. V.; Pereira, S. L.; and Barros, L. N. 2011. System design modification with actions. Advances in Artificial Intelligence–SBIA 2010 31–40. Wu, K.; Yang, Q.; and Jiang, Y. 2005. Arms: Actionrelation modelling system for learning action models. In Proc. ICAPS, volume 5, 2003. Zhang, Y., and Ding, Y. 2008. CTL Model Update for System Modifications. J. Artif. Int. Res. 31:113–155. model M actions A domain MA Figure 4: α-CTL Model Update. In order to extend the CTL model update proposed by Zhang and Ding (2008) for planning, we have also defined a set of primitive operations, named Primitive Update in Action operations (PUAs for short). In (Menezes, Pereira, and Barros 2011), we defined a set of four PAUs to update a plan without modifying the planning domain, that are: PUA1 - adding transitions induced by an action; PUA2 - removing transitions induced by an action; PUA3 - adding a new state; and PUA4 - removing an isolated state. In (Menezes, Pereira, and Barros 2010), we defined three extra primitive update operations to update a planning domain (i.e., to modify action specifications), that are: PUA5 - adding transitions by relaxing an action precondition; PUA6 - adding transitions by adding an action effect; and PUA7 - removing a transition by deleting only one of non-deterministic effects). Formally, given a complete model M∗A induced by a set of actions A; a partial model M ⊆ M∗A ; and an α-CTL formula ϕ defining a planning goal, the set of primitive update operations PUA1-PUA4 can be used to refine and validate the partial model M. If this is not possible, i.e., there is no way to modify M to satisfy ϕ throughout PUA1-PUA4, then the planning domain has to be updated by using operations PUA5-PUA7. Considering all the primitive update operations, PUA1PUA7, we have defined a minimal change criterion that is able to suggest a set of minimal changes to the planning domain designer. Discussion and Future Work The main motivation for this work is to build a tool that can support a knowledge engineer to specify a planning domain for real applications where there is no guarantee to find an optimal or correct plan solution. Therefore, one of the challenges for future work is to define an α-CTL model update algorithm that is scalable with 1842