Average Reward Reinforcement Learning

点击翻译按钮获取中文摘要

IRI-9520243 Prasad Tadepalli Oregon State University $77,772 - 12 mos. Average Reward Reinforcement Learning Reinforcement learning is the improvement of performance of an agent due to rewards and punishments. Most studies of the phenomenon so far have assumed that the learning proceeds through maximizing the discounted total reward from the environment. Discounting tends to favor the sacrifice of higher long-term rewards in favor of short-term rewards. On the other hand, performance is often better enhanced in real situations by the use of average reward per unit time. One such application is the scheduling of automatic guided vehicles (AGVs - robots used for material handling tasks in flexible manufacturing systems), where optimization of discounted total reward will influence the robots to serve machines that are close and neglect those far away, leading to stoppage of the more distant machines, and perhaps to the halting of the entire manufacturing system. The objective of this project is to use the AGV scheduling problem as a means of studying reinforcement learning methods that optimize average reward to avoid such pitfalls, leading to a better understanding of the factors that influence performance under automated average reward reinforcement learning.