title

Are we computationaly happy with Baldwin effect?
A re-consideration on Hiton \& Nowlan's experiment.

propose 
a model to which lamarkism is possible to be implement



A key to understand the intention of their experiment
an example of genotype
(1 1 0 1 0 9 0 9 1 0 9)
its possible phenotype
(1 1 0 1 0 1 0 1 1 0 0)
or 
(1 1 0 1 0 0 0 1 1 0 1)
etc.


Mills & Watson (2005) wrote
"In typical runs of the simulation after very few generations, the all 1fs phenotype is found by lifetime learning, and the average num-ber of 1's in the genotype increases rapidly. In subsequent generations the average number of 1fs in the genotype increases slowly towards the fittest genotype." => ???


"The fitness of each individual is the average fitness of the phenotypes it produces during its lifetime and the fitness of each phenotype is Fmax if it is all 1fs and 1 otherwise." => ???


Mills & Watson (2006) wrote
"We find the extent of modifications to this landscape provided by plasticity, by calculating the
 expected fitness of individuals as a function of their location with respect to the two peaks."

As in the previous year's paper 
"In this model each of the phenotypic trials is independent
and the fitness of an individual is calculated by the mean
of the fitness of all of its phenotypes."


"Our evolving population is modelled using a constant
population size of 200 N-bit genotypes, initialised on the
low-fitness peak. Each new generation is formed by
fitness-proportional reproduction."

Again this makes readers wander how


And concluded
"Thus from an engineering perspective,
the Baldwin effect is not an efficient means for crossing
valleys in terms of the number of fitness evaluations."


whichever assimilation or canalization will be effective
our concern is ...


1 To consider performance from an engineering perspective, we
find the learning population takes approx. 13.3*200*1024 =
2,723,840 evaluations, which is significantly more than for the
non-plastic case (approx. 2382*200 = 476,400 evaluations).
However, this model is not intended to show any engineering
advantage; from a biological viewpoint, generation time is
approximately fixed. Thus the important comparison to make is
upon the number of generations required.


J. M. Baldwin (1896) 
"A New Factor In Evolution."
American Naturalist Vol. 30, pp. 441-457, 536-554.

R. Mills, and R. A. Watson (2005)
"Genetic assimilation and canalisation in the Baldwin effect."
In Proceedings of the European Conference on Artificial Intelligence,
pp. 353-362, Springer-Verlag.