Agents Based on Self-organizing Maps

Perceiving and processing the environment of an agent sensibly can be equally as complex as constructing the agent itself. Not only do many agent implementations require a discrete domain of possible states, but a state as input to an agent should sometimes carry meaningful information instead of just being an abstract number. In such situations, the internal structures of a state space can be exploited—ideally in an adaptive way that maximizes the agent’s performance, even when other agents in a multi-agent system change their behavior. Therefore, two methods were proposed and tested to present information about such structures of state to a surprisingly simple agent model. One of those methods utilized Self-Organizing Maps to achieve a remarkable reduction of state complexity.