After my mutant in Experiment 14 came within 2.5% of GnuBG 2-ply checker and 3-ply cube, as I had mentioned in a recent comment to Ian Shaw in DailyGammon, I wanted to see if I could further simplify my mutant while keeping it achieve same/similar results. I took the previous cube points:
cubepoints = [0.55, 0.60, 0.65, 0.70, 0.75] takepoints = [0.15, 0.20, 0.25, 0.30, 0.35] beaverpoints = [0.39, 0.41, 0.43, 0.44, 0.47] And only changed the beaver points to be also be in 5% increments, like this: beaverpoints = [0.35, 0.40, 0.45, 0.50, 0.55] and set too-good point to 0.85. The first points in the three sets are 0.15, 0.35 and 0.55 in 0.20 increments and the following points are in 0.05 increments. So, all one needs to memorize are 3 numbers: the 0.15 first take point with 0.20 and 0.05 increments. Oh, and the too-good point. Nothing can be simpler. I don't know if the checker play effects the cube play but this time I set it to 3-ply also for both sides anyway. Surprisingly, this simpler mutant performed even better. In games won, it was just as good as GnuBG 3-ply checker + 3-ply cube; in points won, it was just 1.32% shy of it. After 20,000 games bot won 37,854 (51.32%) and mutant won 35,900 (48.68%) of points; bot won 10,003 (50.02%) and mutant won 9,997 (49.99%) of games. Then I ran another experiment to see if more conservative cube points can achieve better results against the same plies of GnuBG. I merely added 0.05 to all of the above numbers: cubepoints = [0.60, 0.65, 0.70, 0.75, 0.80] takepoints = [0.20, 0.25, 0.30, 0.35, 0.40] beaverpoints = [0.40, 0.45, 0.50, 0.55, 0.60] After 20,000 games bot won 23,373 (51.77%) and mutant won 21,779 (48.23%) of points; bot won 9,773 (48.86%) and mutant won 10,227 (51.14%) of games. Mutant won merely 0.55% less points but won 1.12% more games. What's really striking is that the total number of points went down from 73,754 to 45,152 by a whopping 40%. Obviously ppg went down accordingly by 40% from 3.69 to 2.26 It's ineresting to see that when the mutant cubes conservatively, GnuBG dances to his tune and that only 20% less total cube actions, (going down from 15,821 to 12,574), causes twice as much, (40%), difference in total points and ppg's. With these last two experiments, total number of games/trials in my "fartoffski" mutant cube skill experiments reached 140,000 which I believe makes the results significant and conclusive enough, that the "jackoffski cube skill theory" is just a pile of cow pies and that even an arbitrarily fabricated simple cube strategy based on "game stages" and roughly approximate, evenly incremented, round number cube points based on GWC's alone, can perform as good as GnuBG 3-ply (maybe even 4-ply which is not yet tested). To see complete descriptions, stats and analysed results of Experiments 16 and 17 just go directly to: https://montanaonline.net/backgammon/py2.php and scroll down to the bottom. MK
