Dynamical Adaptation of a Neural-Net Based Agent

This paper shows a study of the dynamical adaptation of a neuralnet based agent. We use a recurrent neural net (RNN) scheme which combines prediction and reinforcement learning. We investigated how the internal dynamics of the RNN and the resultant goal-directed behaviors are self-organized while the internal rewarding system of the agent is dynamically changed according to the agent’s own goal-achievements. Our simulation results showed that limit cycling dynamics appear as a solution for the goal-achievement. The whole history of the adaptation process fluctuated by repeating intermittent phase transitions of the internal RNN dynamics from one limit cycling to another with generating diverse goal-directed behaviors.