Self-generation of reward by human interaction — Adaptation to multitask by reflecting hope degree for priority

摘要

In recent years, a robot is required to achieve human need in human living space. Also, complicated and flexible behavior is required for multi tasks in human living space. In such a robot, studies have proceeded by using reinforcement learning. Reinforce learning is highly applicable to the real environment. When using reinforcement learning, it is necessary to design a reward function. The paperproposed self-generation of reward using general indicators for reward function. As a general indicator, we set indicators that mimic the creature's sensory organs. In the indications, a reward is generated based on pleasure and unpleasant in response to sensory input. It is thought that creatures feel unpleasant when input is too strong or too weak and feel pleasant at just right time. It is thought that creatures feel pleasant when prediction of input is easy and feel unpleasant when input prediction is difficult. In the index, pleasant and unpleasant are generated as input with the strength of input and predictability of input. The reward function gives a big reward when feeling pleasant based, and a small reward when feeling unpleasant on this index. It generates external input based on interaction with the environment, and generates reward using this index. Using this indicator for tasks eliminates the need to design reward functions for individual tasks.