Рет қаралды 331
BetaTetris have a strong tendency to play on left-well. The exact reason is still unknown despite my efforts trying to find it out; it just seems extremely hard for BetaTetris to figure out right-well strategies on its own.
Before this, I always added high additional rewards on right-well to force it to play right-well. This time I decided to let it do what it wanted. Fortunately, it did impress me on level 29 playing. The average score of this agent in 20 Hz & no adjustments & 230 lines is 1.16 mil, surpassing StackRabbit 2.0's 1.13 mil (though StackRabbit's number is only an average on 10 games, thus not necessarily accurate). This does not necessarily mean left-well are better at level 29, though, since it may be the case that BetaTetris' left-well playing is better than StackRabbit's right-well playing. (The current best right-well version of BetaTetris only averaged 1.06 mil on this format.) But the difference between left- & right-well may not be that significant in this format.
These two games are the top 2 and the only 1.5M games of ~200 games it played, so the fact that they happened b2b is just by pure luck. The agent didn't have the 230-line limit in mind, so it didn't try to score high at the endgame in the second game.