【摘要】 针对多智能体协同执行围捕任务面临的动态性,提出了一种基于规则与改进wall-following的深度强化学习算法(Rule-Based Deep Reinforcement Learning,RBDRL)。首先,RBDRL算法根据目标和障碍物在历史行为区间内动作选择的统计,进行连续执行多步动作的状态预测,并利用基于wall-following规则设计的Upward-Downward规则在四边形网格环境中生成闭环轨迹;其次,针对闭环轨迹中的冗余路径,采用缩减规则对轨迹进行优化;再次,将这些规则集成到深度强化学习框架中,并设计了综合型奖励机制,尤其在团队奖励中,特别纳入了对时间成本的考量;最后,将RBDRL算法分别与基于计数的深度强化学习算法和无规则的深度强化学习算法在包含不同规模和数量的静态与动态障碍物场景中进行对比实验。实验结果表明,所提方法在解决多智能体在动态环境中协同执行围捕任务的问题时,具有可行性与有效性。更多还原

【Abstract】 In response to the dynamics faced by multi-agent collaboration in performing encirclement tasks,a deep reinforcement learning algorithm based on rules and improved wall-following(Rule-Based Deep Reinforcement Learning,RBDRL) is proposed. Firstly,the RBDRL algorithm performs the state prediction of multi-step actions according to the statistics of the action choices of targets and obstacles in the historical behavior interval and uses the Upward-Downward rule designed based on the wall-following rule to generate a closed-loop trajectory in the quadrilateral-grid environment. Secondly,the trajectory of the redundant paths in the closed-loop trajectory is optimized using the reduction rule. Thirdly,these rules are integrated into the deep reinforcement learning framework,and a comprehensive reward mechanism is designed,especially in the team reward,especially the consideration of time cost. Finally,the RBDRL algorithm is compared with the counting-based deep reinforcement learning algorithm and the rulefree deep reinforcement learning algorithm for experiments in static and dynamic obstacle scenarios containing different sizes and numbers of obstacles,respectively. The experimental results show that the proposed method is feasible and effective in solving the problem of multi-agent cooperative hunting tasks in a dynamic environment更多还原