[computer-go] Example for Morphys law in Evaluation

[Wang Yizao]

Magnus Persson wrote:

> Quoting Chrilly <c.donninger at wavenet.at>:
>
>>> It is late here in Europe and this post is not very coherent. So 
>>> here is the
>>> point: "Static evaluation of groups in go is very very difficult. 
>>> Monte Carlo
>>> evaluation was sent from heaven".
>>>
>> One should always inspect the gifts carefull. Even if the gift  comes 
>> from heaven. It can be a Trojan horse.
>
>
> I see monte carlo evaluation as something that not yet has reached its 
> full
> potential.
>
> With static full board evaluation one needs a complete and correct
> implementation of all possible situations that are likely to arise on 
> a go
> board. If there is limits in the evaluation then I think search will 
> not be
> able to compensate for this.
>
Probably I didn't well understand what you mean by 'search' here. Could 
you explain a little more it?

> With MC evaluation one can add some simple local knowledge which 
> improves the
> evaluation in most situations, and search work quite well in removing 
> those
> moves that evaluated too high early in search.

Do you mean pruning by search here?

>
> I attach a very nice 9x9 test position (from a real game on CGOS) I 
> have used
> during the development of Valkyria as an example of what MC-eval + a 
> litlle
> knowledge can do. White plays first and should try to live but Black 
> has some
> really nice tesujis to keep white from living but white can connect 
> all stones
> and win. There is also a very nice sequence where black kills white if 
> white
> makes the obvious response.
>
> The nice thing about this position is that black can keep attacking 
> white to for
> about 11 ply which Valkyria now can see in a few minutes with very 
> selective
> UCT-search.

What do you mean by selective UCT-search here?

I ran MoGo to solve the problem, and it is always happy to play the white.
The worse is, even I could not see this nice sequence of black that 
kills the white...

MoGo proposed E1 D2 H4 very quickly and I can't see how to kill the 
white after.
What is the tesuji for black then?

>
> When I started working on Valkyria it would sometimes not even get the 
> first
> move right in 10 minutes of thinking time, but with some basic go 
> knowledge it
> now can go very deep. I think this positions is special in the sense 
> that it
> can be searched very deep quickly, but the forcing moves that has to 
> played for
> the first 6-7 ply are not trivial.
>
> If your static evaluation is good then you might not need to search 
> deeper than
> 6-7 ply to see that white is alive or connected, but if the evaluation 
> makes
> errors then this might be really hard for traditional full board search.

Here are you talking about the 19x19 board of the 9x9 one? I think UCT 
works already pretty well for the 9x9 and it is not easy to overpass it 
by adding static evaluations (since it is often not that good).

Yizao

>
> -MP
>
>------------------------------------------------------------------------
>
>(;GM[1]FF[4]SZ[9]AP[SmartGo:1.4]
>PW[b]
>PB[w]
>KM[0.5]
>;B[ge];W[eh];B[ee];W[dd];B[de];W[ce];B[cf];W[bf];B[cd];W[be];B[fe];W[cc]
>;B[bg];W[bd];B[cg];W[ec];B[df];W[hc];B[gc];W[gb];B[hd];W[fc];B[gd];W[ag]
>;B[hb];W[bh];B[ch];W[ci];B[di];W[gh];B[ic];W[hi];B[hg];W[hh];B[fh];W[fg]
>;B[fi])
>
>  
>
>------------------------------------------------------------------------
>
>_______________________________________________
>computer-go mailing list
>computer-go at computer-go.org
>http://www.computer-go.org/mailman/listinfo/computer-go/
>