Presenting deep reinforcement learning algorithm in pursue-evasion problem for the smart police agents

The development and use of artificial intelligence methods to solve different problems have been a vast and active research field for many years. The problem of pursuit-evasion has been used as a testbed in much new research on machine learning and artificial intelligence. In this research, the polices-robbers problem, as a specific form of the pursuit-evasion problem, is considered, in which some police agents pursue a robber agent. The objective of the study is to train two intelligent police agents using deep Q networks so that they can return the robber (escaping agent) to a specific position in the shortest possible time. Two models were presented using the mentioned algorithm in two different scenarios to learn from the experiences of the police agents and finally the performance of the proposed models was tested and evaluated by comparison with the exact brute force algorithm. After training the agents, it was observed that in both scenarios, the cost of the networks gradually decreased and the amount of rewards received by the police agents at the end of the training increased and converged to certain values. It was observed that the police agents are completely successful in returning the robber to the specific position and in the second scenario, in more than 90% of random environments, they perform this operation successfully.