Computational Intelligence, Volume 00, Number 0, 2012 DESIGNING PROTOCOLS FOR ABDUCTIVE HYPOTHESIS REFINEMENT IN DYNAMIC MULTIAGENT ENVIRONMENTS GAUVAIN BOURGNE 1 AND NICOLAS MAUDET 2 1 CRIL—CNRS UMR, Universit´ e d’Artois, Lens, France 2 LIP6, Universit´ e Pierre et Marie Curie, Paris, France This paper studies multiagent systems where each agent has access to local observations of a dynamic environment and needs to build from this partial information an hypothesis on the state of the system. Each agent ensures that its hypothesis is consistent with its observations, and communicates with other agents to refine this hypothesis by confronting them to their own views. However, these communications are restricted by temporal and topological constraints, and can only be bilateral. We first study in this paper an abstract model of this problem, identifying conditions under which satisfying states can (or will) be reached. We rely in particular on a compositional consistency relation. We then detail a case study involving agents able to reason abductively (with Theorist), and study how demanding are the conditions required in this context. Different bilateral protocols are finally introduced and formally studied, to account for both compositional and noncompositional settings. Received 22 June 2009; Revised 16 July 2012; Accepted 16 July 2012 Key words: agents, hypothetical reasoning, communication protocols, abduction. 1. INTRODUCTION We consider a cooperative multiagent system situated in a dynamic world where each agent only perceives partially the state of its environment (e.g., because of limited range of view). Each agent attempts to build hypotheses to explain its observations and predict the global or future state of the environment, because agents may have to individually exploit this hypothesis during the process. To overcome this lack of information, agents may coordinate and refine their hypotheses by confronting them to other agents’ observations. Ideally, each agent will exchange information with others until it gets an hypothesis that is valid for the observations of each of the other agents. If communications are constrained, in the sense that only some agents can exchange information at a given time, it becomes crucial for them to adequately choose (i) which agent to communicate with, and (ii) what information they will (try to) exchange with them. This defines a problem of collective hypothesis refinement under communicational constraints. Potential applications include decentralized situated agents teams designed to collaboratively achieve a global objective, where each agent is equipped with reasoning abilities, is evolving in dynamic environment (hence, communication links may not be stable), and may have to take decisions on the basis of their current working hypothesis . This paper includes an illustrative example which exhibits these features: a team of agents is to escape from a building on fire. Each agent’s best escape plan to an exit will depend on the hypothesis they make on the supposed origin of the fire. When an agent meets a fellow agent, they may exchange information to refine their view of the world. This paper formally investigates this class of problems. Part of this analysis remains at an abstract level, studying conditions under which it is possible or guaranteed that the different hypotheses of the agents populating the system will be “sufficiently similar” (in the sense that we make precise later on) if agents are only allowed to engage in bilateral interactions, but may not have the opportunity to communicate with every single other agent during the Address correspondence to Gauvain Bourgne, CRIL - CNRS UMR, Universit´ e d’Artois, Lens, France; e-mail: bourgne@ cril.univ-artois.fr. Parts of the material in this paper has previously been presented at DALT 2006 and AAMAS 2007. C  2012 Wiley Periodicals, Inc.