Learning to Survive: Increased Learning Rates by Communication in a Multi-agent System

Abstract

In a hostile multi-agent environment, a team of learning agents utilizing a reinforcement-learning paradigm may not learn at a sufficient rate for the agents to adapt and survive. If we can increase the learning rate of the individual agents, they may learn earlier to select more optimal actions and hence increase the chance of survival. Implementing a simple communication structure between the agents and allowing them to communicate their learning events with team members at each time step considerably increases the learning rate. This increased learning rate can significantly improve the team’s chances of survival in the agent’s environment, but alone is not a guarantee of better performance. The type of information communicated also plays a crucial role in determining increased survival chances. This paper reports on some experimental results from simulating two teams of combative agents, one utilizing reinforcement learning as a control paradigm using communication to increase the measured learning rate.