[Computer-go] Results from 19x19 Valkyria/1k H9 vs Valkyria/10k

Erik van der Werf erikvanderwerf at gmail.com
Tue Oct 5 12:17:56 PDT 2010


2010/10/5 Gunnar Farneb├Ąck <gunnar at lysator.liu.se>:
> On 10/05/10 16:59, Erik van der Werf wrote:
>>
>> Lukasz brings up an interesting point. Winrate may not be the ideal
>> statistic for all situations. Maybe the average score (as used in most
>> early work on MC), soft-max, or a median tracker would be better for
>> some situations.
>>
>> Maybe a nice question for the academics:
>>
>> If you were free to keep track of a histogram for all possible scores
>> in each node (so you have everything from winrate at every possible
>> komi to simply the average score), then what would be the optimal
>> selection strategy?
>
> I have long suspected that median score is the way to go.

The median score may be interesting for updating dynamic komi, but for
selecting the best move it is not enough. One reason is that it does
not account for bounds on the distribution (so it can't distinguish a
sure win from an uncertain outcome).


>> And if the above can be answered, what would be the minimum set of
>> statistics needed to maintain similar performance?
>
> But I have never tested to implement it because I haven't found a way
> to compute it without storing the complete histogram, which would make
> the nodes huge.

You can track the median with fixed increments/decrements depending on
whether the score is higher/lower than the current value. The
magnitude of the steps relate to some kind of forgetting factor.
Different quantiles can be obtained by changing the ratio of the
update magnitudes.

Erik



More information about the Computer-go mailing list