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

BGonline.org Forums

Inconsistent XGR++ Evaluations

Posted By: Philippe Michel
Date: Monday, 19 June 2023, at 6:32 p.m.

In Response To: Inconsistent XGR++ Evaluations (Timothy Chow)

the seed for XGR/XGR+/XGR++ is supposed to be a combination of a hash of the position and a fixed user seed.

That would explain the difference mentionned in Nack's initial post between the XG roller evaluations of a position reached with a 21 and the same position reached with a 62. But not the other discrepancies

Evaluations on different computers could be slightly different if XG detects the CPU features and uses optimized functions that could be 8-way vectorized if possible, else 4-way, else scalar. But I would expect these differences to be much smaller than what was shown, probably invisible if you cannot look at more than 4 decimal digits.

Another thing that may be worth trying is to ensure that the rollouts are not multithreaded and see if it makes a difference. Multithreaded rollouts may show the kind of small differences from the above paragraph (for the same reason: addition is not associative on computers), but they shouldn't show larger ones.

there's something peculiar about the Nackgammon starting position.

The evaluation of the equity of the starting position will be inaccurate anyway since XGR doesn't know that the first roll cannot be a doublet.

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.