Fine-grained multiagent systems for the Internet

Abstract

We present an agent architecture and a system structure that support fine grained multiagent problem solving over the Internet. In our model, each agent works on genetic programming tasks, resides at a workstation on the Internet and is capable of coordinating its activities with a variable number of other agents using the same coordination protocol. The system of agents has a virtual topology that is structured as a loosely coupled network of domains, each of which contains some number of nodes in the overall system. Agents (nodes) cooperate via DGPP, a contract net based high level communication and control protocol. A testbed based on the model has been implemented in Java. Several standard problems were tested. The results showed near linear speed up using six Pentium Pro/Windows NT workstations.