To : All Subj: Interesting paper Parisi, Domenico, Stefano Nolfi, and Federico Cecconi. Le
From: Wesley R. Elsberry 12 Jul 94 08:59
To : All
Subj: Interesting paper
Parisi, Domenico, Stefano Nolfi, and Federico Cecconi. Learning,
behavior, and evolution. Technical Report PCIA-91-14, Institute
of Psychology (C.N.R. - Rome).
We present simulations of evolutionary processes operating on
populations of neural networks to show how learning and behavior
can influence evolution within a strictly Darwinian framework.
Learning can accelerate the evolutionary process both (1) when
learning tasks correlate with the fitness criterion, and (2) when
random learning tasks are used. Furthermore, an ability to learn
a task can emerge and be transmitted evolutionarily for both
correlated and uncorrelated tasks. Finally, behavior that allows
the individual to self-select the incoming stimuli can influence
evolution by becoming one of the factors that determine the
observed phenotypic fitness on which selective reproduction is
based. For all the effects demonstrated, we advance a consistent
explanation in terms of a multidimensional weight space for
neural networks, a fitness surface for the evolutionary task, and
a performance surface for the learning task.
Parisi et al. note that learning and behavior tend to get short
shrift when it comes to evolutionary theories. They cite the
index to Mayr's "Growth of biological thought", where only three
entries exist for "behavior", and no entries for "learning" can
Parisi et al. use neural networks and genetic algorithms to
explore some of these neglected issues.
In a population of neural networks whose inputs are angle and
distance to "food" sources and whose outputs control "motor
actions", normal GA techniques result in an increase in
performance of food finding over time.
Then, a learning task is added: the neural networks now also
predict the outcome of motor output, that is, a prediction is
made of what the sensory input at the next time step will be.
No learned change is transmitted to offspring: the untrained
NN representation is what is acted upon by the GA. However,
with the learning task in place, much better food finding
performance emerges faster in the population.
A supplementary learning task which is not evaluated in
determining fitness of the phenotype was added. For this,
Parisi et al. chose the exclusive-or problem. The naive
expectation was that no correlated increase in XOR learning
would be observed, since XOR learning performance was not
evaluated. However, further generations of the NNs did
learn the XOR task better and faster.
All in all, this was a very intriguing paper.
E-Mail Fredric L. Rice / The Skeptic Tank