Greg Detre
16/5/00
stick to GAs � just a search algorithm
search through space to optimise parameters
in real evolution, evolving DNA � protein
in GAs, not maximising offspring so much in a particular task
Travelling Salesman
want to solve a problem, parameters � find good consolation of parameters
np complete � too big
!n�������� n=10, 3.5m
I think Reil used inversions a lot in his Travelling Salesman solution
same problems can't use the BP algorithm, e.g. recurrent
fully recurrent net
GA makes no assumptions about the problem
GAs ≈ 10 times slower than backprop
because backprop knows about its landscape
both use local search
if trapped in local optima
mutation (especially when trapped)��� Anastasoff
evolution has tried to keep mutation as low as possible
GAs = very similar to evolution � local optima
asymptotic graph is always produced (generations on x vs fitness on y)
Gould + Eldridge � punctuated equilibria � jumps in fossil record
Reil � just a property of the landscape (looks like little steps in the progress towards the asymptote)
comes out of local optima within peaks
underlying dynamics of GAs are similar to evolution
mutation = very important in real DNA because recombination affects high-level interactions between proteins
but altering individual proteins requires mutation to affect single amino acids
Holland �much more brute force, recombination equivalent to mutation in very simple system
they just had huge numbers of very simplee chromosomes
Gaussian distribution of mutation
lots of spacer DNA because you don't want recombination to break up genes
can be a good mechanism sometimes
but a lot of problems don't have these building blocks
e.g. in recurrent NNs, all of the parameters depend on each other
then, recombination = a nasty macromutation
inversion + transdozation(???) = affected by gene regulators
in nature, DNA is processed sequentially rather than in one go
in human, 30,000 genes, but genotype � DEVELOPMENT � phenotype
what response is there to be made to Minsky�s claim that GAs aren't even the best search algorithm???
are we keen on GAs partly because they may not be wholly optimal, but they�re very good within a biological space, and moreover, they produce life-like results
what is it about GAs that make them a good search algorithm???
is it possible that the genome codes for different transformations to be performed on different parts of itself??? does it actually program in the transformations at all, or are they designed �accidents�???
MathEngine � physical environment simulation
Natural Motion organism comprised of multiple chromosomes (NNs)
�evolution has tried to keep mutation as low as possible� � is this really strictly true???
ah, he doesn't like mutation � he prefers crossover etc.
spacer = junk DNA???
having all that spacer DNA to prevent genes being broken up by recombination sounds like a huge waste � why not just have end-of-gene markers???