[Carl-Haakon]

I don't really know what to call this, so I'll call it retrodiction. What I mean is, if you have a model that knows the probabilities of certain outcomes based on some prior information, but you can't measure that information directly, retrodicting that information would be making a statistical "prediction" about them based on the outcomes. For instance:

Say we have an extremely simple two-player game where each player can either play aggressively or defensively. Let's also say that we have calculated these probabilities for how the game will play out thusly:

If player A plays aggressively and player B plays passively, A has a 90 % chance off winning
If player A plays aggressively and player B plays agressively, A has a 50 % chance off winning
If player A plays passively and player B plays agressively, A has a 10 % chance off winning
If player A plays passively and player B plays passively, A has a 50 % chance off winning
No ties are possible; A not losing implies A winning and vice versa

However, player strategies are not available online so we cannot know about them in advance. But if we observe the game after it's concluded and we find that A won the game, we could easily guess A's strategy to be aggressive with a probability of 70 percent (given no prior information). If we then assume that the players' strategies don't change, we would be able to update the probability to which we assign A's being aggressive in a bayesian fashion after each game, and thus be able to figure out A's strategy without actually watching the game.

I'm trying to do something like this with a monte carlo horse racing simulator in order to infer jockeys' strategies without having to watch actual races (they can be quite tedious). I was thinking that by randomly assigning variables to the different jockey's strategicality and examine the simulated races that ended in the way that the actual races did, I'd be better able to infer these values of the jockeys. Has anyone tried anything like this? Are there any articles on this specific subject? Does it have a name or is it just a variation of monte carlo simulations?

[matthew919]

Read up on hidden markov models: http://en.wikipedia.org/wiki/Hidden_Markov_model

I know nothing about horses, so I have no opinion on the validity of your approach. But you should be able to find some statistical software that can implement this type of model rather easily. Good luck.

[Carl-Haakon]

Ah that seems to be it! Thanks a ton!