Multiagent reinforcement learning with the partly high-dimensional state space

✍ Scribed by Kazuyuki Fujita; Hiroshi Matsuo

• Publisher: John Wiley and Sons
• Year: 2006
• Tongue: English
• Weight: 342 KB
• Volume: 37
• Category: Article
• ISSN: 0882-1666
• DOI: 10.1002/scj.20526

✦ Synopsis

Abstract

One method of designing a multiagent system is called multiagent reinforcement learning. In multiagent reinforcement learning, an agent also observes the other agents as part of the environment. As a result, as the number of agents increases, the state space increases exponentially (curse of dimensionality), and the learning speed decreases dramatically. The amount of memory required for learning also becomes enormous. Modular Q‐learning, which was proposed as a technique for solving this problem, has the disadvantage that the learning performance decreases due to the incompleteness of perception. In the current research, the authors propose the HMQL technique for improving the learning performance of Modular Q‐learning by using a method of partially increasing the dimensionality of the state space. © 2006 Wiley Periodicals, Inc. Syst Comp Jpn, 37(9): 22–31, 2006; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20526