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

BGonline.org Forums

Rollout pitfall?

Posted By: Timothy Chow
Date: Wednesday, 2 September 2009, at 8:58 p.m.

Lately I've been experimenting with the following procedure. Say there are only three plausible ways to play a roll. I first do the obvious thing: I mark the top three moves and tell gnubg to roll them out and report the results. I next try to get at the equity a different way: for each of the three moves, I edit the position by manually making the move, and then I ask gnubg to roll out the cube decision for the opponent. (Obviously this works only if the cube is in the center or owned by my opponent; let's assume this is the case.)

Theoretically, the equity should be the same in both cases (except for a minus sign). However, in practice, there is a discrepancy in the results reported by gnubg. It is to be expected that there will be small differences, but recently I tried a position and got very different results. In particular, the "normal" rollout method said that Play A was slightly better than Play B, while the "reverse" method said that Play B was much better than Play A. This was with 1296 trials at 2-ply/2-ply. I can post the position if people are interested, but my guess is that this is a rather common occurrence.

My question is, what exactly is going on here? I looked at the recent thread on "Why do bots have an even/odd effect?" which seems related, but that discussion seemed to be focused on evaluations rather than rollouts. A practical question is, what should one do in the situation I described above? Trust one way and not the other? Repeat the rollout with more trials and/or higher-ply evaluations, hoping that the results will eventually converge?

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.