I just ran into the same problem (on a smaller scale) in another context. I
want to evolve a strategy for what's called "combat" in the AI Challenge
<http://aichallenge.org/>contest from last Fall. (It's over now, but I'm
using it in a class.)

The problem is that combat is between two (or more) teams (of ants). How
well a combat strategy does depends on the strategies of the other team(s)
at the time of the evaluation. I want to evolve all strategies against each
other, but it's not clear how to take into account the dependency of a
fitness value on the population in existence at the time it is calculated.
I don't remember if a good approach to that has been developed. I couldn't
think of anything very clever.

*-- Russ Abbott*
*_____________________________________________*
***  Professor, Computer Science*
*  California State University, Los Angeles*

*  Google voice: 747-*999-5105
  Google+: https://plus.google.com/114865618166480775623/
*  vita:  *http://sites.google.com/site/russabbott/
*_____________________________________________*



On Mon, May 21, 2012 at 11:25 AM, Roger Critchlow <[email protected]> wrote:

>
>
> 2012/5/21 Owen Densmore <[email protected]>
>
>> Either of you finish the paper?  Comments?
>>
>>     -- Owen
>>
>>
> No, I can't seem to read anything these days.
>
> But the paper on the neural networks evolving strategies to play the
> prisoners' dilemma with each other was very much a comment.
>
> The fitness of an inherited strategy is defined entirely by the population
> of strategies it is born into, so you cannot evaluate the fitness of a
> neural network independent of the environment in which it plays.   The
> environment in which it plays is determined by the fitness of the
> strategies in play in the last generation of the game and random number
> generators.  The entire system is deterministic, you can integrate from any
> initial state by running the simulation.  You can generate an ensemble of
> outcomes by varying random number seeds and running simulations.  Now,
> having run as many simulations as your budget allows, what do you know
> about the laws governing the system?
>
> You know that the "organisms" evolve larger neural networks even though
> size is penalized, and that you don't have sufficient budget to enumerate
> the possible neural networks, or the possible populations of neural
> networks, or the possible random number streams mutating the outcomes of
> generations, or the possible encounter schedules between members of a
> population (which will matter when the neural nets learn to implement
> reinforcement learning).  Although you know everything about the bits and
> the deterministic rules that make a particular simulation, you don't know
> squat about laws that allow you to predict the outcome of the next
> simulation.  Each generation of simulation is a law unto itself.
>
> -- rec --
>
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org
>
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org

Reply via email to