Hi Yamato,
I like you idea, but why do you use only 5x5 and 6x6 Go?
1. Our second algorithm, two-ply simulation balancing, requires a
training set of two-ply rollouts. Rolling out every position from a
complete two-ply search is very expensive on larger board sizes, so we
would probably have to consider some subset of leaf positions. We
wanted to analyse the full algorithm first, before we started making
approximations.
2. We can generate a lot more data on small boards, to give high
confidence on the results we report.
3. IMO it's important to do the science to understand underlying
principles first, and then scale up to bigger boards, more complex
Monte-Carlo searches, etc.
I don't think the 200+ Elo improvement is so impressive
I agree that it would be much more impressive to report positive
results on larger boards. But perhaps it is already interesting that
tuning the balance of the simulation policy can make a big difference
on small boards? Also, larger boards mean longer simulations, and so
the importance of simulation balancing should become even more
exaggerated.
because the previous approaches were not optimized for such a small
boards.
I'm not sure what you mean here? The supervised learning and
reinforcement learning approaches that we compared against are both
trained on the small boards, i.e. the pattern weights are specifically
optimised for that size of board.
I agree that the handcrafted policy from Fuego was not optimised for
small boards, which is why it performed poorly. But perhaps this is
also interesting, i.e. it suggests that a handcrafted policy for 9x9
Go may be significantly suboptimal when used in 19x19 Go. So
automatically learning a simulation policy may ultimately prove to be
very beneficial.
I'm looking forward to your results on larger boards.
Me too :-)
Coming soon, will let you know how it goes.
But I hope that others will try these ideas too, it's always much
better to compare multiple implementations of the same algorithm.
-Dave
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
