Improve Performance of Attack in the Team Robots Soccer using Reinforcement Learning

Due to the impossibility of predicting all possible states for agents in a wide dynamic multi-agent system, machine learning methods are useful tools to control agent behavior. Simulated Robot Soccer is a well known multi agent benchmark to evaluate machine learning algorithms. In this paper, QV-Learning algorithm (a well known reinforcement learning algorithm) is used to improve the performance of the attack in 2D robots soccer team. The reinforcement signal is defined based on the players involved in the attack can reach the ball in front of goal or lose the ball; They receive positive and negative reward according to the mentioned status, respectively. We use the idea of division the reinforcement signal proportional to the amount of expertness (knowledge) of agents to improve the performance. Here, the expertise is defined as the difference between highest action value and lowest action value in the each state. The simulation results show using the idea of expertise improves the train speed and the performance.