[Computer-go] Computer-go Digest, Vol 12, Issue 79

[Hendrik Baier]

>   Indeed. My plan is to generalize this scheme with more strict 
> conditions. Maybe we even can combine LGRF with the information of 
> RAVE (inspired from Arpad Rimmel's works). If the learning works well, 
> it should fix a lot of errors in my rules of the playout features. 
> This might be a way to make the playouts to learn how to play correct 
> semeai moves.
>
If you know in a playout that the best reply has to be one of only a 
handful of options (e.g. in a semeai), non-zero probabilities for all of 
those moves plus LGRF should be a nice way of making the search adapt. 
But whenever you make the conditions more specific, be aware that you 
will get fewer samples per condition, and you need lots of samples to be 
able to count on good replies staying in the table longer than bad replies.

If you could find an efficient way of combining LGRF with local 
patterns, I would see some potential. I couldn't get it to run fast enough.