Is maximin a machine learning algorithm, or simply another concept which is used for increasing the probability of winning?
migrated from stackoverflow.com Oct 2 '12 at 19:27
This question came from our site for professional and enthusiast programmers.
Machine learning usually uses sets of past data to create patterns and predictions on future data. For instance, manually categorizing previous letters that people have drawn in the training stage in order to detect what letter someone is drawing.
As far as I know, maxmin doesn't have any sort of training based on past data, so wouldn't really fall into machine learning. (This sounds like a homework question though, you might want to check your text book's definition of machine learning to be sure).
Maximin (if it is the same as minimax?) involves creating an objective function based on possible payoffs, i.e. you have the objective of minimizing the maximum possible loss.
I could imagine a classification scenario where there are payoffs associated with each actual target label/prediction label pair, and you use the maximin critieria to build an objective function that is minimized by the learning algorithm..., but I have not seen this before. Usually the payoff structure is +1 for correct 0 for incorrect, which results in maximizing the probability that the predicted label is correct. However, I am sure there are plenty of real world examples where a maximin type of objective would make more sense.
However, I am not sure how many standard machine-learning algorithms are adaptable to correspond to this type of optimization criteria...