file.pdf (175.11 kB)

The Necessity of Average Rewards in Cooperative Multirobot Learning

Download (175.11 kB)
journal contribution
posted on 01.01.2002, 00:00 by Poj Tangamchit, John Dolan, Pradeep Khosla

Learning can be an effective way for robot systems to deal with dynamic environments and changing task conditions. However, popular single-robot learning algorithms based on discounted rewards, such as Q learning, do not achieve cooperation (i.e., purposeful division of labor) when applied to task-level multirobot systems. A task-level system is defined as one performing a mission that is decomposed into subtasks shared among robots. In this paper, we demonstrate the superiority of average-reward-based learning such as the Monte Carlo algorithm for task-level multirobot systems, and suggest an explanation for this superiority.

History

Date

01/01/2002

Usage metrics

Exports