Sender: |
|
Date: |
Thu, 29 Dec 2005 17:58:25 -0500 |
MIME-version: |
1.0 (Apple Message framework v746.2) |
Reply-To: |
|
Content-type: |
text/plain; charset=US-ASCII; delsp=yes; format=flowed |
Subject: |
|
From: |
|
In-Reply-To: |
|
Content-transfer-encoding: |
7bit |
Comments: |
|
Parts/Attachments: |
|
|
>> Liviu is our coevolution man, but I believe this will do the
>> following:
>>
>> For each individual i in the population
>> If the individual has not yet had GROUPSIZE evaluations
>> Pick a random individual j from the population that
>> has not yet had GROUPSIZE evaluations
>> evaluate i and j together once, assigning fitness to
>> each
>>
>> Sometimes it happens that there's no one on left to evaluate against,
>> in which case a random person j is chosen and his fitness is not
>> assigned.
>
> Okay, that would explain some things. Is an individual ever matched
> against itself?
We usually check for that not to happen.
>> During breeding, except for a few GP modifiers which had a bug in
>> them, individuals modified to the next generation should be marked as
>> not evaluated. Individuals which are direct copies will be marked as
>> evaluated. However CompetitiveEvaluator will evaluate individuals no
>> matter how they're marked.
>
> In other words, individuals will always be marked as not evaluated?
Yes.
> If so, how can I set the initial fitness of an unevaluated individual
> just once?
Use problem.preprocessPopulation(state,state.population)
>> Your GoProblem example is odd though: it appears in
>> postprocessPopulation you're just trimming fitnesses to between 0 and
>> 100. Why would you want to do that?
>
> I set the minimum to zero in the belief that fitnesses have to be
> nonnegative. Is this true?
SimpleFitness may have arbitrary values as of the previous version (I
believe, or if not the current version does not complain about that
anymore).
> I set the maximum to 100 so that an individual could not build up a
> huge fitness by winning many early games, then never be knocked down
> by later losses.
I am not sure I follow the reasoning. Even if an individual won
against early competitors in initial generations, that says little
about how good he is against better opponents found at later
generations. I would treat this on a per-generation basis and ignore
information about wins or loses from previous generations.
Again, I encourage you to look into the single-elimination
tournament, which we found to work nicely. There is also a version
of double-elimination-tournament somewhere, though we have not tested
it that much.
Best regards,
Liviu.
|
|
|