Behavioral Diversity in Learning Robot Teams

Tucker Balch
Ph.D. Thesis
College of Computing
Georgia Institute of Technology

Abstract

This work investigates the origins of behavioral diversity in learning robot teams. Behavioral diversity refers to the extent to which agents assume distinct behavioral roles in a group. Most research in multi-robot teams to date has centered on homogeneous systems, with work in heterogeneous groups focused primarily on mechanical and sensor differences between agents. In contrast, this work examines teams of mechanically identical robots. These systems are interesting because they may be homogeneous or heterogeneous depending only on behavior. Behavior is an extremely flexible dimension of heterogeneity in learning teams because the agents converge to hetero- or homogeneous solutions on their own. This research provides new tools for the investigation of behavioral diversity in multi-robot systems and a significant body of results using these tools in simulated and real mobile robot experiments. The experiments specifically investigate the relationship between the reinforcement function used for training and the diversity and performance of the resulting multi-robot teams.

Contributions include:

• a methodology for building and evaluating learning behavior-based robot teams,
• a taxonomy of multi-robot reinforcement functions,
• a classification of multi-robot tasks,
• a quantitative measure of behavioral diversity in multi-robot teams,
• a quantitative measure of behavioral difference between individual robots,
• evaluation procedures for multi-robot task performance, and
• a significant body of experimental results illustrating the use of these tools in three multi-robot task domains.

An online Java applet that demonstrates the calculation of social entropy (diversity) is available here.

The thesis is available in compressed postscript online (see below). Note: the thesis was formatted for double-sided printing. It will print fine on a single-sided printer but the margins will be a little different on odd and even pages.
• All in one file (8.0MB)

• Chapter 1:Introduction (380K)
• Chapter 2:Background and Related Work (607K)
• Chapter 3:Methodology (808K)
• Chapter 4:Task and Reward (350K)
• Chapter 5:Behavioral Difference and Diversity (948K)
• Chapter 6:Diversity in Robot Foraging (3.3MB)
• Chapter 7:Diversity in Robot Soccer (851K)
• Chapter 8:Diversity in Cooperative Movement (1.2MB)
• Chapter 9:Conclusion (164K)
• Appendices and Bibliography (114K)

My thesis is also available as a College of Computing Technical Report (GIT-CC-98-25). You can request that one be sent to you by sending e-mail to Linda Williams lindah@cc.gatech.edu, or tech-reports@cc.gatech.edu. Let me know if you run into any trouble getting a copy.