Moyo Go Studio DevBlog

Dariush has been shouting that it has a better pattern system than Moyo Go, so since I decided to start working on the program again, I had a look at Dariush. It is easy to make critical comments about the look & feel and functionality of that software, but that might be unfair, as I have no idea how long they have been working on it. The pattern system is what I was afraid of - and so I should be:

I took the first game in one of my SGF folders marked “Cyberkiwon 7d”. I have the folders a through z there, with numbered games in them. The first game was numbered 9, and the file is here. The strongest Cyberkiwon games are often played by pro’s, and statistical analysis shows hat most games in that category use pro Joseki.

Of that 126-move game, Dariush predicted 55 moves correctly, Moyo Go 54. You can try it yourself and let us know here if you found a mistake.

This (statistically insignificant) result does not daunt me. On the contrary. Darius obviously has a much less sophisticated pattern system than Moyo Go has (when I look at which info is obviously in its patterns and even the size of its pattern database). So there must be something very important the Dariush dudes have implemented, and Moyo Go didn’t implement that at all. So if I figure out, using some detective work, what that crucial difference is the programmers of Dariush “got”, and I didn’t get, then wham, Moyo Go will be the undisputed leader again.

I am excited because I already think I got it. I think I know what Dariush has, in its pattern system, that Moyo Go hasn’t. It is obvious, when replaying 9.sgf in both programs. (On a tangent: I think Moyo Go actually outperforms Dariush here, because where some moves are not at all considered by Dariush, Moyo Go actually classifies them as the second-best choice, then again I haven’t counted how often that happens with both programs). And Darius, contrary to Moyo Go, does not give statisticaly sound move estimations, but hey, as long as the moves are good :-)

Anyway. I think that Dariush (it seems obvious when you observe its pattern classifications) greatly emphasises locality. Computer Go researchers are divided over whether locality should be used in a move-suggestion system, but Erik van der Werff seems to use it to good effect. And now Dariush too, it seems. If this is really true, if locality plays a role in move prediction, then adding it to Moyo Go should give it a crushing ahead to Dariush.

I have once talked about locality in move prediction with Michael Reiss. We both thought it was “sinful”, bad, unscientific, etc. But first Erik, and now Dariush, hm..

You see, competition is good. One of the first things I’m gonna do is to add locality to Moyo Go’s pattern system, and test it.

I’ll be back.