| |
BGonline.org Forums
Bob Koca Interview--A Must Read
Posted By: Henrik Bukkjaer In Response To: Bob Koca Interview--A Must Read (phil simborg)
Date: Friday, 8 October 2010, at 7:05 a.m.
There are plenty of room for improvement in the bots.
Especially if you'd like them to win the highest possible percentage of actual games, not "just" play the teoritical lowest possible error rate.
An example: Human players find it very easy to spot positions that will lead to high or low error rates for the remainder of the game. IE. if a certain position will lead to "complex" or "simple" checkerplay decisions for one or both sides. That often affect our decisions, or at least it should. Especially if you and/or your opponent is not playing 1,7 ER or thereabouts.
This is a very difficult concept for a bot.
I don't have XG (yet), but to my understanding, it's threaded so that it is capable of analyzing during play (just as humans unconsciously do). This is one important prereq to improving actual bot play vs. human opponents. But such a thing as estimating the complexity of a position, and calculating that into a decision is also important in getting better. I think you'll need a separate neural net trained in assessing this - and I think it's fairly simple to train it (it can be fed positions from analyzed matches recorded online), getting trained at guessing error rates to come, etc. Then incorporating this info into the bot decision making will also require some thinking in another direction than used to.
I really think there's a LOT to gain in these areas for bot play, and thus for analyzing positions from human vs. human matches.
gnubg being able to use skewed ME tables and lopsided rollout noise adjustments, was just scratching the surface - and in my oppinion, it's not being used nearly to the extension that it should be.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.