Teaching a series of actions by the universal evaluations of each sensory information

Various studies related to reinforcement learning(RL) have been performed. RL is a simple and powerful learning method so it is used for a real robot. In ordinary RL, a reward function is designed to teach a given task. Generally the design of this function is difficult and laborious work because of the need of the consideration about a given task and environment beforehand and the need of the expert knowledge about a reward. This means an agent learned from external sources, not autonomously. To solve this problem, we proposed the method of machine learning through interactions which is independent of any task and environment. In previous works for path planning problem, the method can lead an agent to a goal point by the interaction of an agent with a human and environment. But the effects from a human and environment were mixed and we did not confirm that a human can teach an agent a series of actions under the condition that environment forces an agent to take different actions. In this paper, we recruit four men as participants and instruct them not to lead a goal but to teach a route. We experiment with path planning problem and confirm that a human can teach an agent his/her intention perfectly.