A Simplified Pursuit-evasion Game with Reinforcement Learning

Abstract

In this paper we visit the problem of pursuit and evasion and specifically, the collision avoidance during the problem. Two distinct tasks are visited: the first is a scenario when the agents can communicate with each other online, meanwhile in the second scenario they have to only rely on the state information and the knowledge about other agents' actions. We propose a method combining the already existing Minimax-Q and Nash-Q algorithms to provide a solution that can better take the enemy as well as friendly agents' actions into consideration. This combination is a simple weighting of the two algorithms with the Minimax-Q algorithm being based on a linear programming problem.