A Research On Energy-efficient Algorithm For Multi-agent Collaborative Searching In Uneven Terrains Based On Deep Reinforcement Learning

In recent years,as the problems to be solved by agents become more and more complex,scholars have gradually changed from single agent system to multi-agent system.Multi-agent systems are often used in search traversal,unmanned monitoring and other scenarios.There are many mature algorithms using multi-agent system to search an area of interest,but these algorithms are basically designed for flat terrain,and the environment searched and traversed in practical application is often uneven mountainous terrain.In the search algorithm designed for flat terrain,the agent usually uses the shortest path to move between navigation targets.In flat terrain,such motion saves time and energy,but in uneven terrain,such motion often leads to excessive energy consumption.However,the existing agents basically use portable energy to provide power,so the energy consumption of agents should be reduced as much as possible to avoid the lack of energy before the multi-agent system has traversed the region of interest.In order to solve the above shortcomings,two multi-agent search energy-efficient algorithms for uneven terrain environment are proposed.The main work and innovations are as follows:1.Deep reinforcement learning is introduced into multi-agent system.By designing a reasonable deep reinforcement learning reward and punishment function,the next optimal target location to be traversed by each agent is planned in advance,so as to avoid the problem of repeated coverage as far as possible,improve the coverage efficiency and reduce the total energy consumption of the system.2.For the multi-agent system to search the uneven terrain environment,two neural network models are designed,which enable the agent to select the next optimal location to traverse in real time.For other complex environments,we can also refer to these two network models to construct a more suitable network model.3.Combined with the gradient characteristics of uneven terrain,terrain adaptability is introduced,which can guide the agent to take the optimal energy consumption path as far as possible.Terrain adaptability is added to the kinematic model,and an energy-efficient coverage algorithm based on terrain adaptability is designed.4.The intelligent physical energy consumption model is added to the reward and punishment function of deep reinforcement learning,and the next target point to be traversed is selected with the lowest global total energy consumption as the goal,and then an improved energy-efficient algorithm based on the reward and punishment function is designed.Simulation results show that the two algorithms proposed in this thesis have better energy saving than the traditional search traversal algorithm in uneven mountain environment.Moreover,when some agents in the system break down and disconnect,the two algorithms will not affect the continuous work of other agents,and have good robustness.