Title: An Experimental Study for Development of Multi-Objective Deep Q-Network - In Case of Behavior Algorithm for Resident Tracking Robot System

Year of Publication: Dec - 2020
Page Numbers: 7-16
Authors: Masashi Sugimoto, Ryunosuke Uchida, Haruka Matsufuji, Shinji Tsuzuki, Hitoshi Yoshimura, Kentarou Kurashige, Mikio Deguchi
Conference Name: The Sixth International Conference on Electronics and Software Science (ICESS2020)
- Japan

Abstract:


Over the years, many studies have been conducted with the objective of facilitating the working of robots in dynamic environments. Various robots have been developed to assist humans in workspaces, such as a house or factory. In general, robots are required to work effectively and safely in a dynamic environment to achieve their tasks. Inaddition, the robots should recognize state as similar as Human. However, it is not easy to make a robot behave like a human in dynamic environments. In this study, the person watching robot system will be developed. In detail, in this paper, the flexible-analyze system of the robot will be focused on. In this study, it is used that Deep Q-Network (DQN) of convolutional neural network to estimate the Q-value itself, so that it can obtain an approximate function of the Q-values. However, it seems to the following of multitasking and moving goal point tracking that Q-Learning was not good at has been inherited by DQN. In this paper, to upcome the weak points of DQN by changing the exploration ratio as known as " dynamically, has been tried. In addition, the authors had been tuned the network layers of DQN for generalize to the given environment. In the proposed algorithm, a resident-tracking task and an energysaving task are holding. The whole tasks are working based on DQN. Thus, the behavior algorithm of the person watching robot system will be developed. From the verification experiment, the simulation results showed the proposed method has been acquired actions to switch the desired task.