A Mixed-Experience Multi-Agent Reinforcement Learning Method for Motion Planning

Abstract

The invention discloses a mixed-experience multi-agent reinforcement learning motion planning method, that is, ME‑MADDPG algorithm. This method is trained by the MADDPG algorithm. When generating samples, it not only generates experience through exploration and learning, but also increases the high-quality experience of successfully planning multiple UAVs to the target through the artificial potential field method, and stores these two kinds of experience in the Different experience pools. During training, the neural network collects samples from two kinds of experience pools with changing probabilities through dynamic sampling, takes each agent's own state information and environmental information as the input of the neural network, and takes the speed of the multi-agent as the output. At the same time, during the training process, the neural network is slowly updated, and the training of the multi-agent motion planning strategy is stably completed, and finally the multi-agents can autonomously avoid obstacles in a complex environment and reach their respective target positions smoothly. The invention can efficiently train a motion planning strategy with better stability and adaptability in a complex and dynamic environment.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to a multi-agent reinforcement learning motion planning method. Background technique [0002] With the vigorous development of scientific theory and science and technology, multi-agent systems are more and more widely used in people's daily production and life. Driving, etc. In these fields, it is necessary to use multi-agent motion planning technology. The multi-agent motion planning problem is a kind of problem of finding the optimal path set of multiple agents from the starting position to the target position without conflicts. How to make the agent avoid obstacles and other agents efficiently and reach the designated area has become a major research problem. [0003] The motion planning methods currently proposed by researchers can be generally divided into global path planning and local path planning. Although global path planning can efficiently and quickl...

AI Technical Summary

Problems solved by technology

[0004] The artificial potential field method has a simple and effective obstacle avoidance planning strategy, but there are problems of local minima and planning that are difficult to apply to dynamic and complex environments; the MADDPG algorithm does not care about the complexity of the environment and has the characteristics of autonomous learning, but it has convergence difficulties and training problems. The problem of poor stability

Images