TITLE

Robot Behaviour Learning:  Modularity and Conflicts

BY

Amitabha Mukerjee
Center for Robotics
Department of Mechanical Engineering
Indian Institute of Technology
Kanpur, India


ABSTRACT

We investigate the problem of learning in behaviour-based robots,
where each behaviour intermediates directly between the input stimulus
and the output action.  Learning occurs within each behaviour module
itself.  We investigate the possibility that these learned behaviours
may conflict with other pre-existing behaviours, by using notions such
as the {\it modularity} of behaviour systems (relative independence
between behaviours).  We show that
    a) As the behaviour system becomes more powerful, its
	modularity or its ability to co-exist with new behaviours
	decreases.  
    b) The process of learning results in more general
	behaviours which are more likely to result in conflicts.
An important practical result is the development of a simple test by
which a behaviour refinement system can identify potential conflicts
that may have developed while learning.