[computer-go] Experiments with UCT

[Rémi Coulom]

Wang Yizao wrote:
> Rémi Coulom wrote:
> 
>> Wang Yizao wrote:
>>  
>>
>>> No, I mean, essentially UCT is a determinist algorithm, but in our
>>> program we have a stochastic part(for giving a score at each leaf).
>>>    
>>>
>> That sounds strange. I don't understand. Can you explain more?
>>
>> Rémi
>>
> In short, let's suppose that the tree has depth 10, which is not the
> case of Go for most of the time, and every simulation arrives the end
> leaf( at depth 10), where the score given is determinist. In this case
> UCT works pretty well I believe. The question is, in this case will it
> be better than, for exemple, min-max tree search?
> 
> Yizao
> 

I am even more confused now than I was before asking my question :-)

I will try to answer your question. Your question depends on what you
mean by "better than". UCT will not find a move better than min-max,
since min-max will find an optimal move (well, in fact, I believe that
UCT's move is likely to perform better against an imperfect player in
case the current position is theoretically lost). But UCT is likely to
find it much earlier. In the ECML paper, the authors show that UCT
outperforms alpha-beta.

This being said, I still don't understand how your question is related
to mine.

Rémi