| |
BGonline.org Forums
Checker play vs deep hi-lo backgame
Posted By: Karol Szczerek In Response To: Checker play vs deep hi-lo backgame (Jack Mack)
Date: Sunday, 26 March 2017, at 7:53 p.m.
I still think there is a general contradiction, not specific to this position, with trusting a full rollout vs mini rollout, etc.
For whatever my opinion is worth (I'm afraid I don't have enough knowledge nor experience to be a serious partner for such a discussion, but let me respond), I think the bot just lacks vision (for the lack of better word) like humans have. It's just a cold-calculating machine. It's "decisions" are based on a mega-rollout of the starting position. Therefore the lower the level of analysis (from rollout to minirollout, from minirollout to 4-ply, from 4-ply to straight evaluation function (0ply?)), the smaller the data input for its evaluation function. On that basis I think the bigger the rollout, the more precise the output.
What causes the bot's weakness in "deep" and/or other certain type of positions? Undertraining? Overtraining?
Positions that I call "deep", which basically means "many checkers back", often produce so many possible moves with a single roll (imagine 8 checkers spread in the outfield and a 11 to play), that sometimes the bot just doesn't see all possibilities, so its calculations are doomed to be flawed by a bigger margin than in the more common positions, where the options are more limited. So I think a programming/memory limit on one side. On the other - probably undertraining. These positions are rare in human vs human and human vs bot play, and humans steer their play to mimic the bot, so my guess is that the same is true for bot vs bot play, which is how the bot trains, hence undertraining.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.