Comment My 'prediction' for this year... (Score 1) 136
Dan, congratulations on some absolutely brilliant code. You kept me up for an extra hour thinking about your algorithm. Pure brilliance.
However, I wonder if one of your insights (or one of your Jesse Rosenstock's, rather) isn't going to make the next tournament much less exciting. In your explanation of Iocaine, you explain the use of a random hedge, and write,
I include this algorithm as a hedge; if someone is actually able to predict and defeat Iocaine Powder with any regularity, before long the random predictor will be ranked with the highest score (since nobody can defeat random!). At that point, the meta-strategy will ensure that the program "cuts its losses" and starts playing randomly to avoid a devastating loss. (Thanks to Jesse Rosenstock for discovering the necessity of such a hedge.)
Suppose everyone in the next tournament employed a hedge of this type. Wouldn't that dramatically randomize the tournament results? For instance, if one player's prediction algorithm kicked the hell out of anothers, the player with the bad algorithm would switch to random play after a while (let's say, 100 turns), and could conceivably catch up, at which point there might be some fair play until it kicks into random again very soon afterwards. There couldn't be any more huge margins in the scores, like there were between Iocaine and some of its competitors. The margin would be small enough, in fact, that natural chance could throw the weaker algorithm to the top.
It seems to me that the hedge function makes code cheap - if you're losing, you're guaranteed something close to a tie. Bad algorithms will bring the good ones to a middle ground. I may be entirely wrong about this... I'd appreciate it if someone else went over my theory.
Of course, if this turns out to be a problem (I'm not sure it will: I'm no algorithm genius, but even if I'm right, there will probably be many competitors who won't incorporate the hedge) there's a simple solution. Disallow any random or pseudo-random number generation. Make the contest entirely prediction based. This may fly in the spirit of game theory and of the original rules, but it might be necessary in the future.
My first Slashdot post! Boy o boy!
Richard Dub
However, I wonder if one of your insights (or one of your Jesse Rosenstock's, rather) isn't going to make the next tournament much less exciting. In your explanation of Iocaine, you explain the use of a random hedge, and write,
I include this algorithm as a hedge; if someone is actually able to predict and defeat Iocaine Powder with any regularity, before long the random predictor will be ranked with the highest score (since nobody can defeat random!). At that point, the meta-strategy will ensure that the program "cuts its losses" and starts playing randomly to avoid a devastating loss. (Thanks to Jesse Rosenstock for discovering the necessity of such a hedge.)
Suppose everyone in the next tournament employed a hedge of this type. Wouldn't that dramatically randomize the tournament results? For instance, if one player's prediction algorithm kicked the hell out of anothers, the player with the bad algorithm would switch to random play after a while (let's say, 100 turns), and could conceivably catch up, at which point there might be some fair play until it kicks into random again very soon afterwards. There couldn't be any more huge margins in the scores, like there were between Iocaine and some of its competitors. The margin would be small enough, in fact, that natural chance could throw the weaker algorithm to the top.
It seems to me that the hedge function makes code cheap - if you're losing, you're guaranteed something close to a tie. Bad algorithms will bring the good ones to a middle ground. I may be entirely wrong about this... I'd appreciate it if someone else went over my theory.
Of course, if this turns out to be a problem (I'm not sure it will: I'm no algorithm genius, but even if I'm right, there will probably be many competitors who won't incorporate the hedge) there's a simple solution. Disallow any random or pseudo-random number generation. Make the contest entirely prediction based. This may fly in the spirit of game theory and of the original rules, but it might be necessary in the future.
My first Slashdot post! Boy o boy!
Richard Dub
