[Computer-go] Semeais

[Hideki Kato]

Kahn Jonas: <alpine.DEB.1.10.1101151835460.13877 at phare.normalesup.org>:
>On Sat, 15 Jan 2011, Hideki Kato wrote:
>
>> The problems are not only ko fights but also include multiple semeais
>> and discarding plays.
>>
>> In the case of there are multiple semeais on the board, the
>> simulations have to try to solve other semeai because the later can be
>> bigger than the former.
>
>Why? I don't feel that accurate furikawari is needed in the playouts.
>In the tree, yes. But in the MC part?

If the later is bigger and winning the semeai is the only way for 
black to win the game, then (at least) some simulations have to try 
to win the semeai.  Otherwise white wins all simulations.

Ah, my though is in the context of AMAF, which almost all strong 
programs are using.  With AMAF, all potentially good moves have to be 
played in a simulation, to make it possible that the tree part can 
select an optimal move in a few seconds.

>Moreover, when we enter the playouts, as David Fotland
>mentions, all the semeai are "solved" in the sense that one player wins
>it, with one-move margin. If we have automatic replies in all the
>semeai, their status is unchanged throughout the playout.
>
>> Also, discarding stones is a very important
>> technique for strong Go players.  Simmulations have to be able
>> to play other possibilities even if the opponent tried to solve a
>> semeai.
>
>Do simulations need it? For me, that's really a tree-part requirement.
>Playing in a lost semeai (save ko) is a bad move.

Yes, playing in a lost semeai is bad (aside ko threats).  The 
automatic reply, however, forces to win a winning semeai.  In the game 
of Go, there are many positions in which losing a semeai (and play 
elsewhere) is better even if the semeai can be won.  When using AMAF, 
prohibiting such moves is not a good idea.

>Replying is still
>overall in favour of the player who wins the semeai. If we compare to
>the position two moves before, it is better for that player.
>So a mistake in the playout would not be fully punished. The only danger
>I can see in that is if symmetrically they are punished, that is if the
>loser just ignores redundant moves by the winner in the semeai, as
>illustrated by the failure of the conditioning strategy by Olivier
>Teytaud.
>
>
>After all, the aim of the MC part is not to play well; it is to give an
>accurate evaluation of the situation.

In the context of modern MCTS implementations, that's not enough as I 
wrote above.

Hideki
-- 
Hideki Kato <mailto:hideki_katoh at ybb.ne.jp>