Re: Minimization principal for evolution

There is a buzz word in computer science called "Genetic Algorithms",
which try to use evolutionary principles to minimize objective
functions.
The issue is really handicapped in two areas.
1) Evolution does not subscribe to exactly one objective function for
all organisms, wheras all studies in optimization have one objective
function by which all hypothesis are measured. Part of evolution is
the development of different species for different niches (not the best
species for one niche).

I think this is partly because most problems in computer science are
single optimization problems (such as best route for salesmen, best
electric circuit for pre-defined purpose, etc.) rather than arms-race
problems, and partly because the field is just getting started and it
takes time to develop software for arms-race problems compared to
single-optimum problems. But a disadvantage when genetic algorithms are
applied to non-arms-race problems is that the problem is in some sense
irreducibly complex. The target is set way-up-high at the very start,
and the algorithm must get toward that one target in order to show
progress and thereby supply selection pressure. Natural evolution has
it comparatively easier, because at the start the goal (defeating the
enemy) is relatively easy to reach, and it's only after that goal has
been nearly achieved that the goal starts to move, but it moves only a
little bit at a time so it's easy for selection pressure to keep moving
toward its new position, never being so far from the goal that the goal
is invisibly far away.

Here's a problem type where genetic algorithms could make direct use of
arms-race: Simulated terrorists opposed to anti-terrorism forces.
Perhaps our Department of Homeland Security should give it a try.

2) There are much better algorithms for optimization, bootstrapping,
monte carlo, gradient decsent, etc... The gentic algorithms have
problems with actually minimizing a function because one strand will
dominate the gene pool, see issue number 1.

I rather like the thermodynamic (annealing, water-level) approach.
You use a genetic algorithm to do the actual optimizing, but you
contrive a pretend arms-race by raising the water level thereby
reducing the portion of land above water level every so often.

I'm not sure what you mean by "one strand". In a true genetic
algorithm, you aren't evolving a single line of descent, you're running
a whole population of genomes in parallel, and it's only when the count
of some genome drops to zero that you stop emulating it with the
others. Ideally you factor the phenotype somehow so that you can put
different factors on different loci and emulate meiosis to
mix-and-match the various loci/factors stochastially so that several
factors and several combinations of factors can all evolve in parallel.
One bad example is trying to optimize a computer algorithm expressed as
a Von Neumann program, whereby the different instructions in a program
module are so tightly intertwined that mix-and-match (cross-over) is
fatal. But a production-rule algorithm might be more reasonably
factored such that mix-and-match would work just fine. A stochastic
production-rule algorithm, where competing rules would be picked at
random, would work fine, and duplication events would work to increase
the priority (chance of being picked) in case of conflict, so the
emulation would be even more like natural evolution.
..

.

Relevant Pages

  • Re: IDs semantic manipulations
    ... capable of modeling the observed rates of evolution. ... observed site are "random" or a part of some anticipatory algorithm ... computed by the cellular biochemical network. ... This is exactly the same point that the number guessing ...
    (talk.origins)
  • Re: Human design and natural "design"
    ... >> This is also true for some GA's; for example in the checkers playing GA ... >> Abstract-An evolutionary algorithm has taught itself how to play the ... >> suggest that the principles of Darwinian evolution may be usefully ... The GA had a fixed set of built-in fitness functions ...
    (talk.origins)
  • DNA in nonliving systems
    ... Is my tooth brush made out of DNA? ... of the design of either a living or nonliving system. ... non-compartmentalized models of evolution. ... search algorithm dubbed Directed Exhaustive Search. ...
    (sci.bio.evolution)
  • Re: Evidence for Big Leaps?
    ... a more efficient search algorithm is responsible ... for the adaptation. ... the neo-Darwinian search algorithm (random mutation, ... Quite what the relevance of this is to evolution is a mystery to me. ...
    (talk.origins)
  • Re: GAs: linear rather than generational evolution?...
    ... I'd assume - this is a genetic algorithms newsgroup... ... They are certainly used in practice and ... be the "best" algorithm for any ... another issue is mutation rate: ...
    (comp.ai.genetic)