摘要
强化学习是实现多机器人对复杂和不确定环境良好适应性的有效手段,是设计智能系统的核心技术之一.从强化学习的基本思想与理论框架出发,针对局部可观测性、计算复杂度和收敛性等方面的固有难题,围绕学习中的通信、策略协商、信度分配和可解释性等要点,总结了多机器人强化学习的研究进展和存在的问题;介绍了强化学习在机器人路径规划与避障、无人机、机器人足球和多机器人追逃问题中的应用;最后指出了定性强化学习、分形强化学习、信息融合的强化学习等若干多机器人强化学习的前沿方向和发展趋势.
Reinforcement learning （RL） is an effective mean for multi-robot systems to adapt to complex and uncertain environments. It is considered as one of the key technologies in designing intelligent systems. Based on the basic ideas and theoretical framework of reinforcement learning, main challenges such as partial observation, computational complexity and convergence were focused. The state of the art and difficulties were summarized in terms of communication issues, cooperative learning, credit assignment and interpretability. Applications in path planning and obstacle avoidance, unmanned aerial vehicles, robot football, the multi-robot pursuit-evasion problem, etc., were introduced. Finally, the frontier technologies such as qualitative RL, fraetal RL and information fusion RL, were discussed to track its future development.
出处
《西南交通大学学报》
EI
CSCD
北大核心
2014年第6期1032-1044,共13页
Journal of Southwest Jiaotong University
基金
国家自然科学基金资助项目(61075104)
关键词
多机器人系统
强化学习
马尔科夫决策过程
计算复杂度
不确定性
muki-robot systems
reinforcement learning
Markov decision process
computational complexity；uncertainties
