Arimaa Forum (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
Arimaa >> General Discussion >> Harlog vs Winning probability
(Message started by: lightvector on Oct 2nd, 2014, 8:48pm)

Title: Harlog vs Winning probability
Post by lightvector on Oct 2nd, 2014, 8:48pm
Just for fun, I took all rated games from 2008-01 to 2014-06 where both players had a gameroom rating of 2000 or better and looked at the probability that a player won the game conditioned on the material score at that point according to HarLog.

Specifically, for every position in every game that had a winner, I computed the HarLog score in millirabbits from the perspective of the current player (prior to that player's move), and then counted a 1 if that player won that game and a 0 otherwise. I made no attempt to adjust for any biases, such as the bias here towards longer games, or identify any causality, such as material loss being a symptom of a crumbling position as opposed to the cause of it.

Here are the results, bucketed every 500 millirabbits:

Code:
Lower    Upper   Mid     Wins    Total   Prob
-1000000 -9751   -10000  1342    50039   0.0268191
-9750    -9251   -9500   284     4852    0.0585326
-9250    -8751   -9000   237     4547    0.0521223
-8750    -8251   -8500   572     6079    0.0940944
-8250    -7751   -8000   383     6122    0.0625613
-7750    -7251   -7500   617     6888    0.0895761
-7250    -6751   -7000   1037    8342    0.124311
-6750    -6251   -6500   1021    8123    0.125692
-6250    -5751   -6000   1519    11196   0.135673
-5750    -5251   -5500   1629    8443    0.192941
-5250    -4751   -5000   2402    13050   0.184061
-4750    -4251   -4500   2709    13254   0.204391
-4250    -3751   -4000   2513    11947   0.210346
-3750    -3251   -3500   5942    24142   0.246127
-3250    -2751   -3000   4125    15987   0.258022
-2750    -2251   -2500   6397    21642   0.295583
-2250    -1751   -2000   11447   36385   0.314608
-1750    -1251   -1500   11486   30064   0.382052
-1250    -751    -1000   26814   64343   0.416735
-750     -251    -500    10052   21560   0.466234
-250     250     0       220567  438537  0.502961
251      750     500     11936   21368   0.558592
751      1250    1000    37191   61753   0.602254
1251     1750    1500    18191   28391   0.640731
1751     2250    2000    24327   34279   0.709676
2251     2750    2500    14888   20482   0.726882
2751     3250    3000    11703   15236   0.768115
3251     3750    3500    17036   21807   0.781217
3751     4250    4000    9210    11277   0.816707
4251     4750    4500    10101   12405   0.814268
4751     5250    5000    10203   12111   0.842457
5251     5750    5500    6554    7898    0.82983
5751     6250    6000    9158    10345   0.885259
6251     6750    6500    6534    7178    0.910281
6751     7250    7000    7023    7900    0.888987
7251     7750    7500    5939    6422    0.92479
7751     8250    8000    5424    5718    0.948583
8251     8750    8500    5335    5838    0.91384
8751     9250    9000    4078    4255    0.958402
9251     9750    9500    4402    4635    0.94973
9751     1000000 10000   47770   48896   0.976972



Title: Re: Harlog vs Winning probability
Post by Fritzlein on Oct 4th, 2014, 1:01pm
Does that make Harlog more or less predictive than FAME?  One way to compare them would be to find the best-fit conversion of material into winning probability for each (perhaps linear, perhaps logistic?), and see which has the lesser prediction error.



Arimaa Forum » Powered by YaBB 1 Gold - SP 1.3.1!
YaBB © 2000-2003. All Rights Reserved.