[streptomycin]

I'd be pretty curious about that too. I actually tried doing something semi-similar before. My goal was to simulate a basketball game by having every player take an action based on "expected points per possession". So if you are covered tightly but a shooter is open, your expected points would be high for passing but low for shooting. This would then be done recursively for every decision, continuously looking ahead until the shot clock expires (so maybe the best move would be to pass it into the post, draw a double team, and then kick it out for an open shot).

The obvious problem was that there are a ton of free parameters involved and tons of branching. Every little action creates a new branch of possibilities with its own set of probability distributions. So even if (as they describe in the article) helping out more often decreases the likelihood of the current ball handler making a shot, how does that propagate through the rest of the possession?

[abecedarius]

http://en.wikipedia.org/wiki/Markov_decision_process for a start.