|
BGonline.org Forums
Some questions/comments re: opening roll RO's
Posted By: Tom Keith In Response To: Some questions/comments re: opening roll RO's (Tom Keith)
Date: Thursday, 13 December 2012, at 4:38 p.m.
Just exploring the idea of training a human-style bot a little more. Maybe the thing to do is to have the NN estimate the probability that a human player would make a given candidate play. For example, with an opening 3-2 it turns out that humans will play
24/21, 13/11: 51.8% of the time
13/11, 13/10: 40.6% of the time
24/22, 13/10: 5.6% of the timeIf you had a NN that could accurately estimate these numbers, then in the rollout you could randomly choose a play from the candidate list according to its estimated probability.
The probabilities would be different for different strengths of the human player, so it might be useful to train a separate net for strong and weak players. That would let you have a rollout between two strong human-style players, two weaker human-style players, or one stronger and one weaker player. For in-between players you could interpolate between strong and weak.
This system still has defects though. It doesn't account for a human player's game plan. For example, if a player decides it is correct to blitz this turn, he is more likely to continue blitzing next turn. My system doesn't take that into account -- if the decision to blitz is close, the system might randomly choose to blitz one turn and not the next with no regard to an underlying plan.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.