An approach to multi agent pursuit evasion games using reinforcement learning

The game of pursuit-evasion, which is encountered frequently in applications of security, has always been a popular research subject in the field of robotics. Especially in the last two decades, when computer scientists gave rise to learning, the agents turned into intelligent agents and they started to use the information about their enviroment for their own purposes, without using the help of a map. This tendecy drew considerable amount of attention and opened the area to newcomers from several different disciplines. Reinforcement learning, which takes the advantage of an agent's interaction with the environment, is a method widely used in pursuit-evasion domain. With the help of this method, agents use the feedbacks (rewards and punishments) taken from the environment to optimize their behaviour, without using complex sensors and maps. Although there are successful examples of the one-pursuer one-evader scenario, there is not enough research on multi-agent pursuit-evasion problems in literature. In this master's thesis, a research is conducted on multi-agent pursuit-evasion problem using reinforcement learning and the experimental results are submitted. The intelligent agents use Watkins's Q(lambda) learning algorithm for the solution of the problem. Q-learning is an off-policy temporal difference control algorithm. The method we used on the other hand, Watkins' Q(lambda) learning algorithm, is a unified version of Q-learning and eligibility traces. It uses backup information until the first occurence of an exploration. In our work, concurrent learning is adapted for the learning of the pursuit team. In this approach, each member of the team has got its own action-value function and updates its information space independently.