On 9/28/07, Christoph Birk <[EMAIL PROTECTED]> wrote: > > > On Sep 28, 2007, at 4:28 AM, Jason House wrote: > > On 9/28/07, Jason House <[EMAIL PROTECTED]> wrote: > > Since there's obviously some kind of major performance gap, for now > > I'll aim to align with Anchor_1k. From there, I hope it'll be > > easier to diagnose what's going wrong. > > > > Correction: I meant to say GenAnchor_1k > > Hi Jason, > > Looking at the performance of hb-amaf-1k I suspect you have some > serious bug(s)
I agree. I'm actually quite shocked at just how serious. I have unit tests for a lot of the core logic (that run at start up, always ensuring accuracy), and have changed a lot of other logic to conform to what others have done... Those changes even tripped various unit tests and asserts (which then had to change). Before doing all this AMAF testing, I would have sworn that I had a reasonable implementation... Maybe not the best one, but certainly reasonable. Now I know that it's not the case. I think I've pretty much ruled problems with the eye rule, suicide handling (I don't allow it), selection of random empty legal non-eye-filling moves, how the end of the game is detected, and how games are scored. I'm starting to run out of ideas. I'm thinking of doing a million random numbers and seeing if they look uniform, or going back to scour the core logic (that passes all the unit tests). in our code. A simple MC program without ANY heuristics except not > playing iside > your own 1-pt-eyes should get about 1050 ELO (10k simulations, eg. > myCtest-10k). Other configurations of my bot have achieved that rating with 1-ply logic (example: housebot-633-UCB is 1050 ELO).. I haven't tracked the number of playouts (since it's variable), but if they're as fast as the amaf variant, that'd be about 20k simulations per move near the start of the game. Without really knowing what other bots did (such as more than 1-ply), I thought I had a reasonable implementation. I appreciate everyone who's put up simple bots and explained what they did. That has really helped with comparisons.
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
