Which MET to use in XG
Posted By: Rick Janowski In Response To: Which MET to use in XG (Ray Kershaw)
Date: Tuesday, 25 June 2013, at 6:25 a.m.
In Response To: Which MET to use in XG (Ray Kershaw)
The Rockwell-Kazaross MET was a major achievement, obtaining all the match equities for a 15-point match by direct rollout, carried out piecemeal so that all possible resulting scores from any point had already been rolled out, with the MET file in use by the bot suitably updated. Thus, and unlike earler mathematical based METS, the potential for error through invalid assumption was removed. The tabulated values were subsequently extended out to all scores in a 25 point match by extrapolation. The Gnu 11-point match MET, had followed a similar approach but only rolling out to 11-point scores and with weaker rollout settings. (The eXtremeGammon MET which was previously XG's default MET also utilised rollout results, though I don't know the details.) Comparison of the R-K MET with the Gnu MET over an 11-point match shows differences of up to 0.5%, demonstrating that the new MET had made a measurable improvement.
Subsequently the Kazaross-XG2 MET was constructed in a similar manner but using XG2 exclusively as the bot rollout source rather than Gnu. This was rolled out for all scores in a 9-point match, showing remarkable consistency with the previous R-K results. The maximum difference is 0.07% with average difference of 0.02% (treating each difference as a positive result). More than anything else I think this exercise demonstrated the validity of the R-K MET and that future significant improvement is unlikely. Sensibly it seems, the rolling out process didnít continue beyond 9-point match scores. Instead, the Kazaross-XG2 MET directly used the previous results from the R-K MET for higher scores, including the previous extrapolation from 15 to 25 point match scores.
Which is best? There is obviously hardly to distinguish the preference of one over the other, for practical purposes. On the one hand, there might be some tiny improvement in accuracy with the Kazaross-XG2 MET, but this could also just be statistical scatter/noise. On the other hand, the R-K MET has greater consistency, there being only one joint (15 to 16) rather than two with the Kazaross-XG2 MET (9 to 10 and 15 to 16). However, the consistency of results between the two METs indicates very negligible potential effects.
Neil Kazaross and David Rockwell have both made massive contributions to backgammon in their major efforts in tackling and indeed solving the match equity problem (for equally strong opponents at least). I think it is only right that each should be given equal credit. For me personally this is a very good reason to continue using the Rockwell-Kazaross MET.
Messages In This Thread
BGonline.org Forums is maintained by Stick with WebBBS 5.12.