Belief Revision by Multi-Agent Genetic Search

The revision of beliefs is an important general purpose functionality that an agent must exhibit. The agent usually needs to perform this task in cooperation with other agents, because access to knowledge and the knowledge itself are distributed in nature. In this work, we propose a new approach for performing belief revision in a society of logic-based agents, by means of a (distributed) genetic algorithm, where the revisable assumptions of each agent are coded into chromosomes as bit-strings. Each agent by itself locally performs a genetic search in the space of possible revisions of its knowledge, and exchanges genetic information by crossing its revisable chromosomes with those of other agents. We have performed experiments comparing the evolution in beliefs of a single agent informed of the whole of knowledge, to that of a society of agents, each agent accessing only part of the knowledge. In spite that the distribution of knowledge increases the diiculty of the problem, experimental results show that the solutions found in the multi-agent case are comparable in terms of accuracy to those obtained in the single agent case. The genetic algorithm we propose, besides encompassing the Darwinian operators of selection, mutation and crossover, also comprises a Lamarckian operator that mutates the genes in a chromosome as a consequence of the chromosome phenotype's individual experience obtained while solving a belief revision problem. These chro-2 Lamma, Pereira, Riguzzi / Belief Revision by Multi-Agent Genetic Search mosomic mutations are directed by a (logic-based) belief revision procedure that relies on tracing the logical derivations leading to inconsistency of belief, so as to remove these derivations' support on the gene coded assumptions, eeectively by mutating the latter. Because of the use a Lamarckian operator, and following the literature, the genes in these chromosomes that are modiied by the Lamarckian operator are best dubbed \memes", since they code the memory of the experiences of an individual along its lifetime, besides being transmitted to its progeny. Due to the speciic nature of memes, we further gain in accuracy by modifying the Dar-winian crossover operator so that memes of an another agent can be acquired only if they have been checked for consistency, and possibly mutated, as a result of its own observational experience. We believe our method to be important for situations where classical belief revision methods hardly apply: those where environments are non-uniform and time changing. These can be explored by distributed …