Thursday, November 11, 2004

Statistically Significant

From a comment by SirFWALGMan:

why not a writeup about why 5K hands is statistically nothing in Poker

While I can't give a writeup on this myself, much smarter people have, namely some computer science majors at the University of Alberta.

A decent article/research paper on the development of Loki, their computer A.I.-based poker player can be found here: Using Probabilistic Knowledge and Simulation to Play Poker

The gist of the whole thing is that 5,000 hands is a decent sample size if you're playing against a homogeneous set of opponents. That is just not the case at Party Poker. The above paper recommends 25,000 hands as the cutoff for a statistically significant sample size.

My guess would be that these numbers were not derived theoretically using statistical formulas and the like, but more probably developed empirically through computer simulations. With enough simulations, you can easily determine at which point the number of additional samples yields no statistically different results.

Cool stuff if you're into that type of thing.

No comments: