Reasoning about abductive inferences in BDI agents Jo˜ ao Sousa Lopes 1 , Sergio Alvarez-Napagao 1 , Susana Reis 2 , Javier Vazquez-Salceda 1 1 Universitat Politecnica de Catalunya {jsousalopes,salvarez,jvazquez}@lsi.upc.edu 2 Universidade de Vigo sreis@uvigo.es Abstract The capability of a computational system to deal with unexpected, changing situations and limited perception of the environment is becoming more a more relevant, in oder to make systems flexible and more reliable. Multi-agent Systems offer a com- puting paradigm where properties such as auton- omy, adaptability or flexibility are basic in the con- struction of agent-based solutions. However most of current implementations are not flexible enough to cope with important changes in the environment or information loss. In this paper we propose to introduce abductive reasoning mechanisms in BDI agents and show how such agents are able to oper- ate with partial models of the environment. 1 Introduction For over a decade, Software agents have been proposed as a software engineering paradigm which eases the creation of flexible autonomous computational entities specially capable to operate in complex situations. Motivated by the inability of existing manufacturing systems (i) to deal with the evo- lution of products and (ii) to maintain a satisfying perfor- mance outside normal operation, agent-based technology is more and more used on industrial setups. But abnormal sit- uation handling in industrial plants is often a challenging ap- plication area even for agent-based solutions. The main agent paradigm, BDI agents, is based on a mentalistic approach which tends to rely on a supra believe defining its knowledge as complete and consistent, even if there is missing informa- tion or imprecision on the expected observations. In this work we want to investigate ways to improve BDI agents to operate in dynamic domains where information about the environment may be incomplete and agents need to establish some hypothesis in order to unblock a given rea- soning process. Our approach is that agents, when faced with a hypothesis or a new piece of uncertain information, would try to seek an explanation or justification for the new hypoth- esis/information. After doing so, it could incorporate the ex- planation into its epistemic state together with the new infor- mation. We model this strategy through the use of abductive reasoning. This allows us to then investigate the role of ab- ductive inference within a belief revision framework. In this paper we not only cover the incorporation of new information but also the removal of information. Abduction, as opposed to deduction and induction 1 , is based on the inference of φ (explanans 2 ) from knowledge of the rule φ → ψ and the observation ψ (explanandum). This means that abduction is not an analytic form of inference, but rather based on the Affirming the Consequent fallacy. Like induction, abduction is defeasible: the arrival of new obser- vations might invalidate prior abductive inferences. The conditions which define when a fact φ qualifies as a valid abductive explanation for an observed fact ψ, with a background theory Θ, are [Sindlar et al., 2009]: • Θ ∪ φ | = ψ • Θ ∪ φ | = ⊥ • Θ | = ψ • φ | = ψ In this paper we do not focus on the abductive logic expla- nation, as it is based on the work by [Sindlar et al., 2009]. Our main concern is to infer some abductive logic conclu- sions at a lower level, and then construct a knowledge model that can be used for believes and desires reasoning approach to the human knowledge retrieval. The structure of the paper is as follows: in §2 we describe the industrial use case scenario that will be used in the rest of the paper. Then in §3 we propose an architecture for a BDI agent capable of abduction-driven reasoning. In §4 we show a concrete example where abductive reasoning is applied. In §5 we discuss how BDI agents can process the results of abduc- tion at the level of their practical reasoning. In §6 we compare our work to other approaches. Finally in §7 we present our conclusions and advance some of our future lines of work. 2 Use case example: industrial process Nowadays industries are facing a profound change towards a wide-ranging awareness of the environmental impact of their inner processes. In order to identify specific ways to im- prove the production sustainability, industrial engineers more 1 Deduction is based on the modus ponens syllogism ({φ,φ → ψ}| = ψ), while induction is based on the inference of φ → ψ as a rule from the observation of φ followed by ψ 2 Some authors call this explanantia.