Re: On the Genetic Algorithm
--0-1942697918-1126739617=:63477
Content-Type: text/plain; charset=iso-8859-1
Content-Transfer-Encoding: 8bit
DB James <guest at RapidEuphoria.com> wrote:
<snip>
I thought some of you might be interested in my recent investigations on the
genetic algorithm. </snip>
I find this fascinating. Some ideas for improvements spring to mind immediately:
1. Number of steps should be quantified as part of fitness--in the "real world",
a 95% correct answer in 5 steps is extremely often more successful than a 100%
correct answer in 100 steps.
2. Even the fittest individual will eventually die and even the least fit
individual may live for a while.
3. Not every very fit individual reproduces, some very unfit individuals do.
4. Probability of survival and reproduction are however reasonably proportionate
to fitness.
A rough suggestion for using these ideas:
1. Calcalate fitness as (100% minus the fitness you are currently calculating)
divided by the number of steps.
2. Normalize these results so that the fittest individual has a fitness of 99%
and the least fit individual has a fitness of 1%, with other individuals with
proportionate values in between. (A simple linear proportion could be used, but a
standard deviation based measure is probably better though harder to code.)
Pehaps 99% to 1% isn't quite right, perhaps 95% to 5% or somewhere in between
would produce better "evolution".
3. Now for each indivdual, calculate whether that individual dies or reproduces
based on fitness. For example, an individual has 90% fitness. It has a 90% of
survival and a 90% chance of reproduction. So we generate a random number 1 to
100 for survival and if the result is 1 to 90 it lives, if 91-100, it dies. If it
lives, then generate another random number 1 to 100 for reproduction....
I would be quite interested in collaborating on this project.
--0-1942697918-1126739617=:63477
Content-Type: text/html; charset=iso-8859-1
Content-Transfer-Encoding: 8bit
<DIV>
<DIV>
<DIV><BR><BR><B><I>DB James <guest at RapidEuphoria.com></I></B> wrote:
</DIV>
<DIV> </DIV>
<DIV><snip><BR>I thought some of you might be interested in my recent
investigations on the genetic algorithm. </snip><BR></DIV>
<DIV> </DIV>
<DIV>I find this fascinating. Some ideas for improvements spring to mind
immediately:</DIV>
<DIV> </DIV>
<DIV>1. Number of steps should be quantified as part of fitness--in the "real
world", a 95% correct answer in 5 steps is extremely often more successful than a
100% correct answer in 100 steps. </DIV>
<DIV> </DIV>
<DIV>2. Even the fittest individual will eventually die and even the least fit
individual may live for a while. </DIV>
<DIV> </DIV>
<DIV>3. Not every very fit individual reproduces, some very unfit individuals
do.</DIV>
<DIV> </DIV>
<DIV>4. Probability of survival and reproduction are however reasonably
proportionate to fitness.</DIV>
<DIV> </DIV>
<DIV> </DIV>
<DIV>A rough suggestion for using these ideas:</DIV>
<DIV> </DIV>
<DIV>1. Calcalate fitness as (100% minus the fitness you are currently
calculating) divided by the number of steps.</DIV>
<DIV> </DIV>
<DIV>2. Normalize these results so that the fittest individual has a fitness of
99% and the least fit individual has a fitness of 1%, with other individuals with
proportionate values in between. (A simple linear proportion could be used, but a
standard deviation based measure is probably better though harder to code.)
Pehaps 99% to 1% isn't quite right, perhaps 95% to 5% or somewhere in between
would produce better "evolution".</DIV>
<DIV> </DIV>
<DIV>3. Now for each indivdual, calculate whether that individual dies or
reproduces based on fitness. For example, an individual has 90% fitness. It has a
90% of survival and a 90% chance of reproduction. So we generate a random number
1 to 100 for survival and if the result is 1 to 90 it lives, if 91-100, it dies.
If it lives, then generate another random number 1 to 100 for
reproduction....</DIV>
<DIV> </DIV>
<DIV>I would be quite interested in collaborating on this
project.</DIV></DIV></DIV>
--0-1942697918-1126739617=:63477--
|
Not Categorized, Please Help
|
|
