Path Planning in Unknown Environment with Kernel Smoothing and Reinforcement Learning for Multi-Agent Systems

In unknown environment, path planning of multi-agent systems is difficult. The popular methods for the path planning, such as reinforcement learning (RL), do not work for these two cases: unknown environment and multi-agent. In this paper, we use a special intelligent method, kernel smoothing, to estimate the unknown environment, and combine it with the reinforcement learning technique. The advantage of the combination of the reinforcement learning and the kernel smoothing technique is we do not need to repeat RL for the unvisited state. The path planning process has three stages: 1) the reinforcement learning is applied to generate the training samples; 2) the model is trained by the kernel smoothing method; 3) the trained model gives an approximate action to agents. Experiment results show the proposed algorithm can generate desired paths in the unknown environment for multiple agents.