Tè
Darwiniano
Lunedì 18
Marzo 2002, Elio Tuci - psicologo che in passato ha svolto il tirocinio al GRAL
- ci ha parlato del lavoro che sta svolgendo come PhD student presso la
School of Cognitive and Computing Sciences, University
of Sussex, UK). In particolare, ci
ha presentato:
Evolving integrated
controllers for autonomous learning robots using dynamic
neural network
Abstract
A growing amount of research in Evolutionary Robotics has been
focusing on the evolution of controllers with the ability to modify the
behaviour of a robot, in order to adapt to variation in its operating
conditions. During my talk, I'll briefly review this research area, focusing
more extensively on the Yamauchi and Beer 1994 work.
Yamauchi and Beer's approach clearly represents one of the first attempts in
which continuous time recurrent neural networks have been exploited to
integrate reactive, sequential and learned behaviour in a simulated robot.
However, I point out that, although a promising one, the Yamauchi and Beer's
approach is not completely satisfactory because the autonomy of the simulated
agent is severely compromised by an external reinforcement signal and
externally imposed modularization of the controller.
Our simulation, go beyond the Yamauchi and Beer approach, demonstrating that
it is possible to evolve an integrated
dynamic neural network, with fixed connection weights and leaky-integrator
neurons that successfully control the behaviour of a khepera robot engaged in
a learning task, similar to one proposed
by Yamauchi and Beer. I show experimental results, and make comparisons with
the original Yamauchi and Beer experiment. I discuss the differences that made
it possible for us to evolve an integrated dynamic neural network where we
assume neither an a priori separated controller module nor an a priori
external reinforcement signal as in Yamauchi94. I end by drawing the work
together in the conclusion.
Reference
copyright ©2002 Raffaele
Calabretta. All Rights Reserved.