Improving Reinforcement Learning Algorithm Using Emotions in a Multi-agent System

A new approach for learning is presented here. The system that is named Sepanta consists of a set of agents that are doing a task. Behaviors of agents are adapted according to emotional signals provided by two parts called emotional critic: one global, generating signal for all agents and, one local, for each agent generating signal specifically for it. The main learning algorithm is Q-Learning that is improved by using these signals. Simulation is done for the task of pushing a mass by a number of robots. The main idea for this work has been a learning method that is tuned by emotion signals supplied by critics for assessing the present situation.