Moyo Go Studio DevBlog

In a previous posting I mentioned Dariush’ pattern system, and how it was on a par with Moyo Go’s in terms of pro-prediction.

(But Moyo Go gives accurate statistical data on moves, Dariush doesn’t and there are many more big differences that make Dariush, in my opinion, not a serious competitor YET).

Of course I have many ideas on how to improve the pattern system, but I never bothered because it was the only pattern system in the world. This has changed and it’s time to do something about it.

It’s not neccessary to modify the pattern database itself or anything -the only thing that has to be improved is the “value” of the patterns. MGS uses 16,777,216 patterns (no idea how Dariush manages to squeeze a pattern in a single bit of memory, MGS uses 8 bytes). If it’s true that Dariush uses 60 million patterns and MGS almost four times less, then there is enormous improvement potential for the latter, as with almost 4 times more patterns, Dariush only gets about the same results as Moyo Go.

The key to improving MGS’s pattern system is to go back to my initial “learning” method. My initial learning method is the one that Microsoft copied after studying my previous blog and website. It’s called “Bayesian Learning” or something like that. The mathematics seem to be complex if one reads their paper, but I intuitively re-invented it and have no clue how the maths work. It provided great results but it took ages to finish, so I devised a very much faster method (hundreds of times faster), which yielded very good results too. I never compared the results of both methods but the Bayesian method, given enough time, will surely produce better results. I discarded it not only for its extreme slowness (months of number crunching), but also because it destroys the statistical data, harvested from the game records.

It seems, however, that not many care much about accurate statistical move percentages, and that a higher pro-prediction percentage might be nicer to have than accurate statistics. Dariush does not provide meaningful data on suggested moves (they appear to fake a percentage), and because they have about as strong a pro-prediction as MGS, I’ll see if I can outdo them, abandoning the statistical part of the pattern system. I could easily provide two versions, one with exact statistics and one without, but with better pro-prediction.

This is next on my to-do list, so I expect to report back about progress in a while. It will take a long time to crunch the data in a Bayesian way, but I expect the pattern system to get quite a bit stronger, especially because I have now tens of thousands of extra pro games to work with. The cool part is that I’ll include the Bayesian-cruncher as a part of the retail version of Moyo Go Studio, so that people can keep optimizing their own pattern database. (Because it may be that years of number crunching is required to achieve the optimum).