Q-Learning Based Search of a Ground Target in a grid with partial information

This paper investigates an optimal strategy for capturing a ground target moving in a grid network. The network is embedded with unattended ground sensors (UGS) to record the target's timestamped information. When the pursuer moves over the UGS, the timestamped information of the target, if available, is instantaneously transmitted to it. We propose a Q-learning-based reinforcement learning algorithm. The pursuer, the learning agent, is trained to make its own decision based on the reward policy defined as a pursuit-evasion game scenario. The purser learns the optimal strategy through both exploration and exploitation in the several iterations of the game. We show that a target moving in a random path with a constant speed, the Epsilon-Greedy Q-learning based algorithm can capture the target before exiting from the grid network.