[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Two things that need to be done with XBG

Posted By: Rich Munitz
Date: Monday, 29 June 2009, at 9:17 p.m.

In Response To: Two things that need to be done with XBG (Stick)

The goal is to determine how the two neural nets compare in strength. XBG's ability to produce GNU's results is not by itself a valid test, since GNU could be wrong.

Neil is right that a head-to-head long duration dueller type match-up (if possible) is the best test of relative playing strength. Ultimately what would be most revealing and most efficient would be a match up using the basic neural net without look-ahead (0-ply for GNU). This is not only the fastest to run, but ultimately the goal is to assess the NN strength and not the lookahead which may not always be apples to apples. And probably money games are the best test, and cubeless money games even better. IMO it is best to focus on how the NNs compare to each other and eliminate the other variables like METs, cube adjustments, etc.

Similarly, you can't just take two rollouts of "normal positions" and judge one better than the other (though it can help to instill confidence that it is trustworthy). What is more useful is to find cases where the bot's evaluations disagree and see which bot's evaluations the rollouts tend to uphold. And of course, the answer to that could be dependent upon position type. There is not necessarily a stronger bot in all situations and it is useful to know strengths and weaknesses. Another test is how close does each bot's eval tend to come to the rollout results.

Messages In This Thread

 

Post Response

Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:

 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.