Propagation and refinement of hypotheses under communicational constraints
Directed by Suzanne Pinson, with the support of Nicolas Maudet
october 2004.
march, 14th, 2008.
Distributed reasoning, Multi Agent Systems, Hypothesis formation (abduction, induction), Multiagent Learning (Concept learning), Communication protocols, non-monotonous reasoning.
This work deals with hypotheses refinement between agents, under communicationnal constraint. Our aim is to enable a network of agents to build from the different observations of its agents through constrainted communication one or more hypotheses that are consistent with the whole set of information of the system. These agents have some certain knowledge, a common theory and some factual observations. Different levels of consistency modelize the adequation of an agent's hypothesis with respect to the observations of other agents. Several protocols and strategies are given to deal with this problem in a general framework, that can be applied to different reasoning mode, such as induction or abduction. Different types of dynamicity are studied. In static settings (no change in system structure or observations during update), global protocol plan all the communications to ensure that the agent find some consistent hypothesis, whereas in dynamic settings (in each turn, links between agent and observations can change), different local exchange are used, in conjunction with dialog initiation protocols, to progressively refine the hypotheses of the agents. These mechanisms have been tested in different settings through several test applications whose results are given.