-
Notifications
You must be signed in to change notification settings - Fork 224
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Feature request: Allow user to adjust MaxPlayouts and numSearchThreads #213
Comments
I high dan AI is indeed requested often enough. I think a better option is an AI option which does a quick extra query, that way analysis is not affected. If you have a table of settings (block, playouts) to strength, that would help. |
I used the first one in the past, put katago has gotten stronger since then. I really didn't care about precise accuracy. Obviously the policy is much stronger than 1 dan! I just wanted the 9d setting to easily beat zen7 9 dan as both white and black. |
I'm not sure threading makes a lot of difference at a few playouts, the locking should make it stronger if anything (explore different moves to 1 visit rather than pile them on top policy). |
You might be right, but it seems like only a few users would need the high dan settings. I don't think it's possble to increase the playouts just 1 at a time if you're using multiple threads. You can really leverage how strong the 40b gets so quickly by letting those few players try to see how far they can go defeating Katago in a step-wise manner, since almost no one will get past 9. If they need analysis, they can just play against the full katago setting max visits to whatever point they have difficulty defeating. Since this high dan opponent is kind of a niche setting, once players reach around 4 dan they can improve alot by playing the full strength katago anyways using the settings already in the program. Sorry for being so wordy, definitely feel free to ignore. I'm not knowlegable enough to see how anything based on the policy can get much past the maximum strength on the policy. But thanks for taking your precious time in this discussion! |
If you do a search, you visit the next node by choosing the max Utility + c * Policy where c varies with total visits and so on. Firing off a bunch of separate queries to all 4-8 top policy moves just gets the evaluation 1 move deeper, which could help if the top policy move is a mistake. I'll have to think about it a bit. |
In case its of interest to anyone, golaxy now has an english site with ranked bots 10k - 9d+. the katago 15b net with 4 playouts (1thread) easily beats the golaxy 4d bot. here is an order to play full strength katago from ddk to 4d+ fox by net size and max playouts. This is the closest I could get to your initial comment about corellating playouts with rank. It's not accurate, but really interesting to see the wide range of strength (almost all of amateur human play) possible with so little computing power and so few playouts. |
Just want to comment that these values can be set in analysis_config.cfg, which if nothing else you can download the default of and then point katrain to use it from within the app, no? |
This issue is stale because it has been open 90 days with no activity. Remove stale label or comment or this will be closed in 10 days. |
This issue was closed because it has been stalled for 30 days with no activity. |
Maybe in an advanced tab under settings, would it be worth it to allow the user to adjust MaxPlayouts and numSearchThreads? By setting numSearchThreads to 1, the user has fairly accurate objective control of Katago's strength at levels above the policy. For example, with the 40b net, MaxPlayouts = 9 , numSearchThreads = 1 will yeild an AI slightly stronger than zen7 9d. Regardless of hardware used, the strength of this bot will be the same. With the 20b net, the same can be said of MaxPlayouts = 16 (maybe 12 I forgot), etc. Alternatively, placing the config file in a directory where the end user can edit it if desired would acheive the same result, since only intermediate and advanced users would do that. Thanks for considering!
The text was updated successfully, but these errors were encountered: