I re-ran the comparison between the normal and symmetric networks, after fixing both their respective training so that they properly calculate board values assuming the opponent holds the dice.
This time the comparison is stark: the symmetric network does much, much worse than the regular one. Here is a plot showing the comparison results. Three series: the blue data show the performance of the normal network playing against the symmetric one; the green line shows the normal network against the benchmark; and the orange line shows the symmetric network against the benchmark.
The y-axis shows probability of win, since again, these networks have only one output. The benchmark is the same one I'm comparing the other new network against (80-node normal network with win and gammon outputs, but not trained efficiently - still, much better than pub eval).
The symmetric network does much, much worse than the regular one. So we can really put this to bed now.
This time the comparison is stark: the symmetric network does much, much worse than the regular one. Here is a plot showing the comparison results. Three series: the blue data show the performance of the normal network playing against the symmetric one; the green line shows the normal network against the benchmark; and the orange line shows the symmetric network against the benchmark.
The symmetric network does much, much worse than the regular one. So we can really put this to bed now.
No comments:
Post a Comment