Escaping local optima in multi-agent oriented constraint satisfaction.
MetadataShow full item record
We present a multi-agent approach to constraint satisfaction where feedback and reinforcement are used in order to avoid local optima and, consequently, to improve the overall solution. Our approach, FeReRA, is based on the fact that an agent’s local best performance does not necessarily contribute to the system’s best performance. Thus, agents may be rewarded for improving the system’s performance and penalised for not contributing towards a better solution. Hence, agents may be forced to choose sub-optimal moves when they reach a specified penalty threshold as a consequence of their lack of contribution towards a better overall solution. This may allow other agents to choose better moves and, therefore, to improve the overall performance of the system. FeReRA is tested against its predecessor, ERA, and a comparative evaluation of both approaches is presented.