W. G. Pritchard Lab Seminar - 109 Boucke Building

	     **October 31, 2001**
	
Distributed Robot Team Learning in Inherently Cooperative Tasks.

Lynne E. Parker, Oak Ridge National Laboratory

Abstract:

In this talk, I will first briefly overview the distributed robotics
research activities at Oak Ridge National Laboratory, showing several
videos of our robotics research.  Then, I will focus more
specifically on our research in the development of learning
methodologies for multi-robot cooperation.   Our multi-robot learning
is focused on the challenging domain of inherently cooperative
tasks.  The application that we study in this context is the
Cooperative Multi-Robot Observation of Multiple Moving Targets
(CMOMMT) problem, which we originally defined.  I will first define
the CMOMMT task and then present results from a hand-generated
(non-learning) solution that we developed.  I will present two
approaches we have developed for learning in this domain, which are:
(1) a reinforcement-learning based approach and (2) a neural-network
based approach.   These results will be compared to the non-learning
solution and to simple naive control solutions.  The learning results
show significant improvements over naive control solutions, with one
of the approaches nearing the performance of the non-learning
approach.  The ultimate goal of this research is the development of
general techniques for learning in multi-robot systems that extend
beyond the CMOMMT domain presented in this talk.