|
Use of the knowledge of Perceptual State Transition in Reinforcement Learning
|
|
Kentarou Kurashige1,Yoshiki Miyazaki2
*1, Muroran Institute of Technology, Email : kentarou.academia@gmail.com
2, NEC Software Hokkaido, Ltd., Email : mydks.public@gmail.com
|
|
Abstract
.Reinforcement learning(RL) is one of machine learning and is often used for actual robot. A reward which indicates a task for robot is most important information on RL. But a way to get a reward will change in multiple tasks. It makes performance of RL to be worse. To overcome this problem, we focus on environmental information. The knowledge on RL is based on reward, so it will be useless when given task changes. The opposite to it, environmental information is independent to reward. It does not chanse and is useful after given task changed and new task was given. In this paper, we suggest the method to acquire environmental state transition through sensors as perceptual state transition and to use it to adapt for new task quickly. In addition to that, we suggest that the method updates its stored knowledge every sensing. We consider the method adapt for change of environment by this update. Using proposed method, a robot can learn effectively for multiple task and dynamic environment. In this paper, we perform two kinds of experiments with simulation. One is for multiple tasks under a static environment and another is for a task under a dynamic environment. We will show the validity of proposed system by these simulations.
|
|
Keywords
:
reinforcement learning; multi-tasks; environmental knowledge; dynamic environment
|
|
URL: http://dx.doi.org/10.7321/jscse.v3.n2.1
|
|
|
|