|
|
Re: thought-experiment
Len Piotrowski (lpiotrow@magnus.acs.ohio-state.edu)
Fri, 20 Sep 1996 20:01:44 GMT
In article <3241F808.7F8F@megafauna.com> Stephen Barnard <steve@megafauna.com> writes:
>[SNIP]
>I'm curious, Lenny. Have you ever heard of the Baldwin Effect?
"
In 1896, James Mark Baldwin proposed that individual learning can explain evolutionary phenomena that
appear to require Lamarckian inheritance of acquired characteristics [1]. The ability of individuals to
learn can guide the evolutionary process. In effect, learning smoothes the fitness landscape, thus
facilitating evolution [2]. This first aspect of the Baldwin effect has recently received much attention,
especially for its applications to computational problem solving [3]. In evolutionary algorithms, local
search is analogous to individual learning. Improvements found via local search change the fitness of an
individual without changing the actual genotype.
Baldwin further proposed that abilities that initially require learning are eventually replaced by the
evolution of genetically determined systems that do not require learning. Thus learned behaviours may
become instinctive behaviours in subsequent generations, without appealing to Lamarckian inheritance
[4]. This aspect of the Baldwin effect deserves more attention. Recent work suggests that intuitive
abilities in human language, physics, biology, and arithmetic may be largely instinctive [5, 6]. The Baldwin
effect can help us understand the relationship between learning and instinct. Furthermore, increased
understanding of this second aspect of the Baldwin effect may enable us to improve the performance of
computational problem solving by hybrids of genetic algorithms and local search algorithms.
A special issue of Evolutionary Computation is planned for 1996, the 100th anniversary of Baldwin's
paper [1]. Evolutionary Computation provides an international forum for facilitating and enhancing the
exchange of information among researchers involved in both the theoretical and practical aspects of
computational systems of an evolutionary nature. Papers are solicited that address both theoretical and
computational work related to evolution, learning, instinct and the Baldwin Effect. Examples of topics of
interest include:
When is learning advantageous? Learning can facilitate evolution by allowing individuals to more
quickly adapt to fitness landscapes that would otherwise be difficult to exploit; however, the ability
to learn weakens the selective forces acting on an individual, which can slow evolutionary change
[7].
When is instinctive behaviour advantageous? Instincts can be fast and dependable in a static
environment, although they may not be able to cope with radical environmental change. Thus the
trade off between instinctual and learned responses may have implications for search and learning
in dynamic computational environments.
From a design perspective, what parts of an individual's cognitive machinery should be modifiable
by experience (local search) and what parts should be determined by evolution (genetic search)?
For example, in a typical hybrid of a genetic algorithm and a neural network, the genetic algorithm
determines the network architecture and back propagation determines the network weights [8]. Is
there a Baldwinian justification for this division of labour?
How does learned behaviour become instinctive? It seems plausible that, for some learned
behaviours, there is no evolutionary path that leads to an instinctive replacement for the behaviour.
For computational problem solving with hybrid genetic algorithms, what techniques can we use to
encourage learned behaviours to evolve into instinctive behaviours?
References
1.J.M. Baldwin, 1896, "A new factor in evolution", American Naturalist, 30, 441-451.
2.G.E. Hinton and S.J. Nowlan, 1987, "How learning can guide evolution", Complex Systems, 1,
495-502.
3.R.K. Belew and M. Mitchell, (editors), in press, Adaptive Individuals in Evolving Populations:
Models and Algorithms, Massachusetts: Addison-Wesley.
4.R.W. Anderson, 1995, "Genetic mechanisms underlying the Baldwin effect are evident in natural
antibodies", in Evolutionary Programming IV: The Edited Proceedings of the Fourth Annual
Conference on Evolutionary Programming, edited by J.R. McDonnell, R.G. Reynolds, and
D.B. Fogel, pp. 547-563. Cambridge, MA: MIT Press.
5.S. Pinker, 1994, The Language Instinct: How the Mind Creates Language, New York:
William Morrow and Co.
6.J.H. Barkow, L. Cosmides, and J. Tooby, (editors), 1992, The Adapted Mind: Evolutionary
Psychology and the Generation of Culture, New York: Oxford University Press.
7.R.W. Anderson, 1995, "Learning and evolution: A quantitative genetics approach", Journal of
Theoretical Biology, 175, 89-101.
8.D. Whitley and F. Gruau, 1993, "Adding learning to the cellular development of neural networks",
Evolutionary Computation, 3, 213-233.
"
See:
http://ai.iit.nrc.ca/baldwin/cfp.html
Excerpts appropos the problem of "learning" for functional adpatational
explanations of culture and human behavior:
"From a design perspective, what parts of an individual's cognitive machinery
should be modifiable by experience (local search) and what parts should be
determined by evolution (genetic search)?"
"Improvements found via local search change the fitness of an
individual without changing the actual genotype."
"Learning can facilitate evolution by allowing individuals to more
quickly adapt to fitness landscapes that would otherwise be difficult to
exploit; however, the ability to learn weakens the selective forces acting
on an individual, which can slow evolutionary change."
"Instincts can be fast and dependable in a static environment, although they
may not be able to cope with radical environmental change. Thus the trade
off between instinctual and learned responses..."
"It seems plausible that, for some learned behaviours, there is no
evolutionary path that leads to an instinctive replacement for the behaviour."
Cheers,
--Lenny__
"If you can't remember what mnemonic means, you've got a problem."
- perlstyle
|