Modulating reinforcement-learning parameters using agent emotions

Paper i proceeding, 2012

An actor-critic reinforcement-learning algorithm using a radial-basis-function network for approximation of the actor and the critic was run on a small-scale multi-agent system with an initially unpredictably hostile environment. The performance of two approaches was compared: having fixed learning parameters, and using modulated parameters that were allowed to deviate from their base values depending on the simulated emotional state of the agent. The latter approach was shown to give marginally better performance once the distracting hostile elements were removed from the environment. This seems to indicate that emotion-modulated learning may lead to somewhat closer approximation of the optimal policy in a difficult environment, by focusing learning on more useful input and avoiding pursuing suboptimal strategies.