edwardyu wrote:I have tried to optimize the piece values of my xiangqi engines. A preliminary result is as follows :-
Code: Select all
MAX All Weighted
vp 23 win 171
vpr 96 draws 62
vb 52 losses 183
ve 51 games 416
vn 197 95% UCB 0.520105
vc 203 win rate 0.488152 0.869043
vr 462 95% LCB 0.427812
win rate 0.000388618
The win rate 0.488152 is under the All column, and we should look for it to improving and converge, is this correct?. There is also a win rate 0.869043 under the weighted column. What does this weighted win rate mean?
Also there is a win rate on the MAX tab : 0.000388618 which is very small. Why?
416 games for tuning 7 parameters are not enough.
"All" column is the win rate over all games so far.
Weighted column is the win rate of samples multiplied by their weights.
Probably the regression is overfitting. I know the current version of CLOP has this problem, sometimes. I will improve it for next version.
If you wish to adjust the prior, on line 39 of src/real/CExperimentFromSettings.cpp, you can change:
pfq.SetPriorStrength(1e-3);
to
pfq.SetPriorStrength(1e-2);
or even
pfq.SetPriorStrength(1e-1);
This should solve the overfitting problem.
I will work on this problem for the next version.
Playing more games will help, too.
Rémi