Cooperative Learning over Composite Search Spaces: Experiences with a Multi-Agent Design System

No. 1: Agents, AI in Art and Entertainment, Knowledge Representation, and Learning

Volume

Proceedings of the AAAI Conference on Artificial Intelligence, 13

Multiagent Learning

We suggest the use of two learning techniques - short term and long term - to enhance search efficiency in a multi-agent design system by letting the agents learn about non-local requirements on the local search process. The first technique allows an agent to accumulate and apply constraining information about global problem solving, gathered as a result of agent communication, to further problem solving within the same problem instance. The second technique is used to classify problem instances and appropriately index and retrieve constraining information to apply to new problem instances. These techniques will be presented within the context of a multi-agent parametric-design application called STEAM. We show that learning conclusively improves solution quality and processing-time results.