[computer-go] hybrid monte carlo / TD neural network position evaluator approach

[Andrés Domínguez]

2006/8/22, George Dahl <george.dahl at gmail.com>:
> The hope would be that the neural net would make fewer samples be
> required and that this would outweight the speed penalty.  If the
> neural net made the monte carlo 100 times slower but made a simngle
> sample 200 times more valuable that would be a wonderful result.
> Unlike the example of playing in 2-space single eyes the nets
> knowledge would be of a more hollistic form that would be created in a
> way very theoretically similar to a monte carlo system (temporal
> difference learning) that would hopefully mesh well with monte carlo
> sampling.
>
> Yes, I expect a very fast neural network very different from the sort
> of thing used in neuro go would end up being the best.
>
> - George

Your ideas sound great to me. Computer Go is very interesting because
everyone make very bad programs. Don't worry if you don't get what you
want, but please try it. :-) Good luck.

Andrés