![]() i used the baduk ai kyu rank which is based on sanderland's calibrated ai mode equation, but there are a few extra levels, like -4 to -8 which are omitted from katrain because they dont increase strength that much with the 20b and lower nets. It is the same underlying equation, but different setting in mobile and on katrain. it is equivalent of hypothetically setting calibrated rank to 5 dan Sorry to not make clear, but the -4 setting is only available in mobile, not in katrain. My strongest recommendation for people without internet who have to play bots alot is to play over a real board with nice stones! thats so much more fun than staring at a computer for an hour or a few days But to give you an idea how strong katago is, the 6b net with 1 playout beat gnugo default (I think level 10) 12-0. ![]() I think gnugo level 1 is too strong for you right now, but could be wrong. but i had trouble registering until i used my real 6 letter first name as my username, capitalizing the first letter. i really like golaxy, which as gotten me alot of downvotes lol. You have to register on their desktop site for English. another beginner friendly app is golaxy, which plays probably from 16k - 6d free but needs internet. But they are fairly strong at their lowest levels. The best go programs to play against if price is not a factor are zen7 and zen6. another alternative is the crazystone pro app. I would also reccommend doing one set of badukpop (app) tsumego everyday for 5 minutes to keep getting better and beating better opponents. i would really recommend that you play the games on a real board and input your moves into the app. Here is how to adjustthe settings in baduk ai. Hopefully you can beat the first bot within a week and start moving up. ![]() Im slowly trying to compile a spreadsheet on how to play against katago and the sanderbaduk based bots on android for someone just learning the rules of go to strong amateurs. The android app baduk ai nicely tries to copies sanderbaduk's formula in his kyu level bots. I think for a really good learning bot, it would take some deliberate style transfer AI that would study different kyu and dan level styles from recorded human games, in addition to the general game playing network. At kyu level, the most humanlike bot I found so far is CrazyStone, available as a free app on Android and iOS. In general, I don't think it's worth it for solo play, because it tenukis all over the board almost every turn. Maybe there were a few other small things, I don't remember. "_enable_ownership": ai_strategy in ,Īdded "ai2gtp.py 14k" as an engine in Sabaki settings (see bot_strategies enumeration in settings.py). ![]() I had katago installed via homebrew (brew install katago).Ĭloned and edited the configuration a bit in ai2gtp.py:.That would be cool, to make it more human-like. I think I saw some ideas on how to reduce it in some issue thread. I guess this randomness is inevitable because of your current move selection algorithm. The 14k configuration is extremely tenuki-happy, but does seem to work at the advertised strength. Do you have any ideas for possible ways to tune down katago in standard GTP mode (beyond visits, which is not enough as you pointed out)? Or would you consider releasing a GTP wrapper that just implements your rank-calibrated mode? I mainly use Sabaki on my Mac, because its UI works best for me.Įdit: people seem to run rank-calibrated katrain bots on OGS, so perhaps that mode is accessible via GTP already somehow? Sorry if I am missing something obvious, but I cannot find the documentation for that.Įdit2: Ah, I guess this is the place to start looking: I just wish there were some easy setup instructions in Readme for using with a client like Sabaki.Įdit3: I managed to set it up locally. Thanks for your work on Katrain, it is a really great idea! You seem to use katago in analyze mode in Katrain, and select sub-optimal moves at your app level.
0 Comments
Leave a Reply. |