Monday, January 29, Volen 101, 2:00-3.00 pm
I suggest a way to make evolved robots capable of dealing with sources of
change that were not included in the evolutionary training.
The method consists of evolving mechanisms of parameter adaptation
instead of conventional evolution of the parameters themselves.
I provide a series of experimental results on physical robots indicating that:
a) The method can develop more complex abilities than evolution alone
b) The dynamics of evolved learning systems are qualitatively different
from conventional evolved systems;
c) The approach scales up to large control architectures;
d) Evolved robots remain adaptive after evolution to several sources
of changes (sensors, environment layout, transfer from
simulation to hardware and across different robotic platforms)
e) When evolving genetic expression, evolution selects expression
of parameter adaptation rather than expression of the parameters.
Host: Jordan Pollack