Refining Action Theories through Abductive Logic Programming Renwei Li 1 , Luis Moniz Pereira 1 , and Veronica Dahl 2 1 Center for Artificial Intelligence (CENTRIA) Department of Computer Science Universidade Nova de Lisboa 2825 Monte de Caparica, Portugal {renwei,lmp}@di.fct.unl.pt 2 School of Computing Science Simon Fraser University Burnaby, B.C. V5A 1S6, Canada veronica@cs.sfu.ca Abstract. Reasoning about actions and changes often starts with an action theory which is then used for planning, prediction or explanation. In practice it is sometimes not simple to give an immediately available action theory. In this paper we will present an abductive methodology for describing action domains. We start with an action theory which is not complete, i.e., has more than one model. Then, after some tests are done, we can abduce a complete action theory. Technically, we use a high level action language to describe incomplete domains and tests. Then, we present a translation from domain descriptions to abductive logic programs. Using tests, we then abductively refine an original do- main description to a new one which is closer to the domain in reality. The translation has been shown to be both sound and complete. The result of this paper can be used not only for refinement of domain de- scriptions but also for abductive planning, prediction and explanation. The methodology presented in this paper has been implemented by an abductive logic programming system. 1 Introduction When reasoning about actions and changes, we often assume that an action theory has been given and described in a formal language or in a framework, e.g. situation calculus [15], event calculus [10], action description languages A [7] and ADL [16], the fluent-features framework (FFF) [19], and their variants or extensions. But little work has been reported on how to obtain an action theory. Assume that we want to generate a plan to make the world in a definite state (goal), but we are not certain about the initial state and the effects of available actions. For example, let’s consider Vladimir Lifschitz’ challenge problem 1 : 1 Vladimir Lifschitz’s email message to lmp@di.fct.unl.pt and renwei@di.fct.unl.pt on March 25, 1996. J. Dix, L. Moniz Pereira, and T.C. Przymusinski (Eds.): LPKR’97, LNAI 1471, pp. 123–138, 1998. c  Springer-Verlag Berlin Heidelberg 1998