Pursuit-evasion game with SARSA learned pursuer

Abstract

Pursuit-evasion algorithms have become more important to domains in robotics, like surveillance with robots. Most of these domains lack a complete model or enough data to work with, reinforcement learning could be a good solution to this. This thesis will show that "a pursuer with a Sarsa algorithm is able to catch an evader in a discrete grid environment.” First, pursuit-evasion games and Sarsa algorithms will be explained. The algorithm used in this thesis implements a simple two-dimensional environment with obstacles in which a pursuer agent tries to catch an evader agent. In the experiments pursuer uses a Sarsa algorithm with different parameter settings against a evader that used two different behaviours. This algorithm will be described further in the method section. The pursuer manages to learn to catch the evader in each tested scenario. There is not a very noticeable difference between the different parameter values used for Sarsa, and there is a difference between the two behaviours of the evader. Finally, there will be discussed what would be interesting for future research. The code used for this thesis can be accessed here: https://github.com/tts118/sarsa-pursuit-evasion .