Is it possible and feasible for an aspireing NN newbie (hopefully not for long) to make agood chess player with NNs? Any information would be helpfull. Thanks!

In my opinion, it is pretty much a lost cause.

To "raise" a NN chess player, you need a learning base, which is basically a set of pairs { given_situation, best_action }, where best_action is the best action to do in the given_situation.

And that's just one obstacle.

There's also that in dynamic learning mode (while actually playing), the artificial player will only be able to compute the feedback on the neural nodes once the game is over, and will thus affect all the actions he took negatively (if he lost) or positively (if he won).

Well, I just think NN isn't the good way.

I guess Reinforcement Learning would provide better approaches (see Q-learning and other reinforcement learning algorithms).

Good luck anyway.

I'm not saying that a neural network approach would be poor, but here is a overview paper that talks about alternative methods that you might find interesting: linky

Quote: Original post by darkpanda
There's also that in dynamic learning mode (while actually playing), the artificial player will only be able to compute the feedback on the neural nodes once the game is over, and will thus affect all the actions he took negatively (if he lost) or positively (if he won).

This is traditionally known as the credit assignment problems. I found an interesting book last year (which is on my desk at work now) on delay learning in neural architectures. There are architectures capable of delay learning... as to whether they'd be suitable for chess, I have my doubts. I tend to think that the work in case based reasoning (associated with plan debugging for learning) will prove most successful in the long term.

Cheers,

Timkin