A Study of Controller Design Using an On-line Evolutionary Reinforcement Learning
Recently, much attention has been focused on an on-line learning such as Reinforcement Learning(RL), because it is expected that artificial systems, especially robots, can cope with their situated environments without human intervention as possible.Based on the above, a number of adaptation mechanisms have been proposed. However, most of them could not work well in real worlds, since the validities of those methods had been demonstrated only in computer simulations. To realize adaptation mechanisms which can be useful in real worlds, it should be taken adaptability and computational efficiencies into account.In the paper, we propose an on-line evolutionary reinforcement learning (on-line ERL), in which Actor-Cntic reinforcement learning and evolutionary recruitment strategy are combined. Although the evolutionary computation (EC) is generally an off-line adaptation, in the proposed on-line ERL, an asynchronous/incremental fitness evaluation method settled the problem.The proposed method has been applied to a task of peg-pushing robot control. Some simulation results show the validity of the proposed method.
Reinforcement Learning NGRBF network Evolutionary Computation Peg-pushing Robot
Toshiyuki Kondo Koji Ito
Interdisciplinary Graduate School of Science and Engineering Tokyo Institute of Technology 4259 Nagatsuta, Midori, Yokohama 226-8502, JAPAN
国际会议
8th International Conference on Neural Information Processing(ICONIP 2001)(第八届国际神经信息处理大会)
上海
英文
703-708
2001-11-14（万方平台首次上网日期，不代表论文的发表时间）