| |
BGonline.org Forums
Kamikaze plays
Posted By: Timothy Chow In Response To: Christian Munk-Chirstensen (Paul Weaver)
Date: Saturday, 13 November 2010, at 1:42 a.m.
Paul Weaver wrote:
roll out some kamikaze plays and post them on this site. That will give you a lot more useful info than asking clueless people like me.
Bot rollouts aren't always trustworthy, and the kinds of positions that I think Christian has in mind are cases in point. If the bot doesn't know how to play correctly with 15 checkers back, then it will naturally shy away from kamikaze plays that result in getting 15 checkers sent back.
Even if you do trust the rollout, positions like that take a very long time to roll out. Each trial will take a long time to complete, and the variance will be large so you'll need a lot of trials. No problem with those with easy access to supercomputers, but for the rest of us, tying up our computing resources doing long rollouts of dubious reliability is not an attractive prospect.
Having said all that, I think that if I were to undertake a systematic study of kamikaze plays, I'd proceed as follows.
1. I'd stick to DMP. This will cut down on the variance and eliminate the problem of gross cube errors.
2. I'd start with positions with relatively few checkers back, where the game is nearing the end, and where the position is relatively easy for a human to understand. This will allow you to do "sanity checks" on the bot and allow you to build up your understanding incrementally. It will also speed up the rollouts.
3. Once I felt I had a good feeling for what was going on, I'd "scroll back" the game position to a slightly earlier state to explore more complicated positions.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.