Exploiting value statistics for similar continuing tasks
Fumihide Tanaka, M. Yamamura · 2004
In this paper, we try to consider interaction design for adaptation from the viewpoint of transfer of knowledge. Advancements in robotics are amazing, and their interaction processes with outside world (including human) are getting to be longer in time scale. We investigate these matters in an abstract agent that faces multiple learning tasks within its lifetime, transferring past learning experiences to improve its performance. We formulize the multitask reinforcement learning problem at first, and then we present two ways of incorporating past learning experiences into the agent's learning algorithm.