基于MATD3算法的多智能体避碰控制  

Multi-agent Collision Avoidance Control Based on MATD3 Algorithm

作　　者：郭雷 梁成庆 GUO Lei;LIANG Chengqing(College of Science,Hohai University,Nanjing,Jiangsu 211100,China;School of Artificial Intelligence and Automation,Hohai University,Nanjing,Jiangsu 211100,China)

机构地区：[1]河海大学理学院,江苏南京211100 [2]河海大学人工智能与自动化学院,江苏南京211100

出　　处：《计算技术与自动化》2024年第1期9-15,共7页Computing Technology and Automation

基　　金：国家自然科学基金资助项目(61976084)。

摘　　要：使用多智能体双延迟深度确定性策略梯度(Multi-agent Twin Delayed Deep Deterministic Policy Gradient,MATD3)算法研究了多无人机的避障和到达目标点问题,首先,利用MATD3算法的优越性提高训练效率。其次,基于人工势场法的思想设计了稠密碰撞奖励函数,使得智能体在没有找到最优解决方案时也能得到积极的反馈,加快学习速度。最后,在仿真实验阶段,通过设计的三组对比实验和泛化实验验证了算法的有效性。Multi-agent twin delayed deep deterministic policy gradient(MATD3)algorithm is used to study the obstacle avoidance and target reaching of UAVs.Firstly,the advantage of MATD3 algorithm is used to improve the training efficiency.Secondly,a dense reward function is designed based on the idea of artificial potential field,which can accelerate the learning speed and help the agent get positive feedback when the optimal solution is not found.In the experiment,the effectiveness of the algorithm is verified by the comparison experiment and generalization experiment.

关 键 词：多智能体 强化学习 人工势场法 避障 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]