Re: AI Methods in Go

[Patricia Hughes and David Elsdon <babel17@xxxxxxxxxx>]

Antti Huima wrote:

> Some people are [again] applying Temporal Difference Learning --- more
> generally unsupervised reinforcement learning --- and function
> approximation methods, i.e. neural network methods, to learning go. It has
> been noted in the literature that often appropriate preprocessing of data
> is necessary for an NN method to work well. I think this is very true in
> go. On the surface, go positions are so chaotic that a raw neural net
> trained via TDL is not expected to learn good go strategies in reasonable
> time.

Agreed. The point is that any technique that is going to learn to play Go in
any reasonable length of time is going to need all the help it can get. Which
is why my explicit coding of Go knowledge is a necessary step. My program
already has some interesting abstractions of the board like influence,
strings, groups, boundaries and connection points of various types. All of
this information can be made available to a learning technique and will
considerably enhance its learning speed.

Regards

David