Agents that comprise a multi-agent system can be either homogeneous or heterogeneous. Heterogeneous teams are often used for complex tasks because they allow agents to be specialized for sub-tasks. However, heterogeneous teams of sub-task specialists are brittle: if one specialist fails then the whole team may fail at its task. An alternative approach is to use a team of homogeneous agents, each capable of adopting any role required by the team’s task, and capable of switching roles to optimize the team’s performance in its current context. We call such a multi-agent architecture an Adaptive Team of Agents (ATA). An ATA is a homogeneous team that self-organizes a division of labor in situ so that it behaves as if it were a heterogeneous team. It changes the division dynamically as conditions change, and if composed of autonomous agents it must be able to organize the necessary divisions of labor without direction from a human operator. In this paper we explore the Adaptive Team of Agents experimentally, using genetic algorithms to train artificial neural networks (ANN) as the “brains” for the agents in a game, and find that it is possible to evolve an ATA with ANN controllers for a simple but non-trivial strategy game.

PDF

In H. A. Abbass and J. Scholz and D. J. Reid, editors, Foundations of Trusted Autonomy, 87-114, New York, 2018. Springer.

@incollection{bryant:trustedautonomy18, title={A Neuroevolutionary Approach to Adaptive Multi-agent Teams}, author={Bobby D. Bryant and Risto Miikkulainen}, booktitle={Foundations of Trusted Autonomy}, editor={H. A. Abbass and J. Scholz and D. J. Reid}, address={New York}, publisher={Springer}, pages={87-114}, url="http://nn.cs.utexas.edu/?bryant:trustedautonomy18", year={2018} }