Hi Magnus, I played a lot of games with integer komi on 7x7 in 2006. I found that very selective search methods fail, because they are happy with jigo, and won't ever search a winning move that happened to look bad after a few simulations. With jigo, you must explore more.
Also, in the early days of Crazy Stone, I was scoring territory instead of winning rate. The problem was even bigger. Rémi On 31 mars 2011, at 17:42, [email protected] wrote: > I did not want to spend a lot of time changing Valkyria completely for > integer komi so I made a very simple hack to cope with the fact that I am > using integers and booleans for all computations of wins/losses in Valkyria. > The hack is that every time a jigo occurs in a playout I randomize which > player should win which means the change is very minor to the program. Thus I > d not to model jigo more than in a statiscal sense. > > I am doing some very deep searches to see what could be good moves for the > most important lines in the opening book. I am using the 7.5 komi book > because it is the most developed, but it may be that a 5.5 book is better if > black is favored by 7.0 komi. > > Anyway I just looked at a deep search spending 10hours of search, where old > branches of the tree is cut off every time the program runs out of memory. > (Which may also be a part of the problem). It turned out that search stuck to > a PV and played to the end over and over again with considering other > alternatives. Valkyria uses exploration = 0. The game it played over and over > was 80 ply deep thanks to a small kofight and all move looks reasonable > although sometimes some moves look a little odd. It does search alternatives > a little bit in each position but it seems to quickly lock onto something > giving 50% for sure. > > My question: Is this normal? Without Jigo this would not be possible because > a deep variation to the absolute end of the game would be a clear loss for > one side and then as the score goes down to 0 the search will explore many > alternatives. > > Could it be that playing alternative moves in this search were all too risky > so that search converges on a very long but guaranteed risk free jigo? > > I see this as a problem because if this happens a lot the program seems not > really to search all options probably. It could be that my program has a bug > or twoo that causes this problem, so therefor I am curious if anyone could > reflect on it or have some experience. > > I believe chess programs uses something call "contempt" to avoid drawing too > much (I think espeically against weaker opponents). > > Best > Magnus > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
