Quote:
Originally Posted by dunningrb
(1) One approach would be to silently simulate the remainder of the game with varying decisions for both teams to determine the best choice for the current situation--like what a very basic chess engine would do.
.
|
I doubt we'll ever see this because it would require a huge amount of processing (number of options AI has times number of reactions human has times number of trials) but this does bring up some super interesting baseball and programming questions.
On the baseball side, how different would the AI be than a real life manager? What things would it do way more often, or never? Think about football: If you did this in a football game teams would go for two and go for it on 4th down far more often than in real life. In a baseball game would they steal more often? Bunt more? How different would pitcher substitutions look?
On the programming side, the two hardest things I can see would be long term effects of decisions and factoring in reactions. It might be that the best thing to do to win a game is to put in a new reliever, but if your bullpen is already exhausted maybe you want to save your last guy to be fresh the next day. That's tough to code around.
Reactions are the biggest hurdle (besides processing time) though. How does the AI factor in my reaction when they make a lefty/righty switch? Maybe 90% of my options are to not react, but in reality I'm going to make the reactive switch, so now for example the AI has a weaker pitcher on the mound still at a handedness disadvantage. Guessing what the other manager will do is part of the gut instinct that makes some managers great. Now, for sure, the current AI system isn't great at this either, but I could see the human manager exploiting the AI in some situations if it were probability based.