Memory-based reinforcement learning of inspect/correct tasks

Document Type

Conference Proceeding

Publication Date

1-1-2020

Abstract

In this paper, we investigate the application of Reinforcement Learning (RL) in robotic inspect/correct tasks such as identifying and fixing rivet failures in airplane maintenance procedures or automated cleaning of surgical instruments in a hospital. The location and form of defects is stochastically distributed and may have substantial variability. In RL, rewards and penalties are defined to help guide the robot toward learning an optimal plan or control policy for a task. RL algorithms explore the solution space to improve performance. RL has been widely applied in robotics and autonomous agents research, but primarily for problems with relatively low variability compared to the task requirements overall. In prior work, we have shown standard RL methods perform poorly on inspect/correct tasks [1]. We develop a new memory based RL algorithm to work with tasks having high variability. Performance is tested in a virtual inspect/correct grid environment and compared against one of the widely used standard RL algorithm’s performance, the SARSA algorithm, with no memory. The results show that the new memory based RL algorithm reduces the number of steps needed to accomplish the task in high variability cases compared to the standard RL algorithm.

Publication Source (Journal or Book title)

Proceedings of the 2020 IISE Annual Conference

First Page

489

Last Page

494

This document is currently not available here.

Share

COinS