|
Post by Steve Draper on Dec 19, 2013 13:01:36 GMT -8
A draw in Qyshinsu (termination after 50 moves without either side winning) scores 0 for both. This seems a little harsh - surely forcing a draw should be better than losing?
|
|
|
Post by alandau on Dec 19, 2013 20:06:47 GMT -8
I agree that this could be a problem. In particular, there's no incentive for a player to prevent the opponent from winning vs. drawing (unlike if the goal value for a draw were 1/1, even). This would be bad in a tournament setting, in particular.
Historically it's been a bit of a pain to get games upgraded for issues like these. Tiltyard does have the technology, thanks to versioning for games that can allow the different versions to keep their own historical match records.
It would be interesting to see how much more common draws end up being if the goal values are switched to 50/50. Players are currently encouraged to make riskier moves in the hopes of an unlikely victory with the 0/0 configuration.
|
|
|
Post by Steve Draper on Dec 20, 2013 5:20:38 GMT -8
Yes, when I was analysing my win rates at this game vs an older version of my own player I found some interesting trends, which relate to use of a different score function.
The older version of my player assumes everything is (to be treated as) fixed sum (i.e. - it hates to lose, even with a high score). It used the score function <my score>-max(<opponent scores> (normalized back to a 0-100 range).
The newer version uses much less of an I-want-to-win bias in order to better handle semi-cooperative games like Alerquerque. It uses <my score>+<win bonus>, where <win bonus> is a small artificial plus for having a score equal to (small plus) or greater than (slightly larger plus) the max opponent score (i.e. - it still doesn't like to lose, but it will swallow its pride for a high goal score).
Interestingly switching between these score functions impacted the win rate in Qyshinsu (not hugely, but significantly), with the newer score function tending to (perversely!) find the draw state more often. This is likely because the older method did not mind 'draws' as much (they scored exactly mid way between a win and a loss), whereas the new method sees the draw as nearly as bad as the loss. Depending on how wins and losses are clustered in state space I suspect this could lead to either better or worse performance.
Given the interesting discriminatory effect the current not-in-the-middledness of a draw has, perhaps it should be retained as an interesting outlier case after all.
|
|
|
Post by Sam Schreiber on Dec 20, 2013 10:31:59 GMT -8
This is just a bug. I'd definitely be happy to deploy a fix for it. Feel free to submit a pull request on github.com/ggp-org/ggp-repository/ with a new version of Qyshinsu that fixes the rules.
|
|