Machine learning ensembles sounds just like monte-carlo tree search (MCTS) techniques (also called UCT), which are used in computer go (and more and more other AI problems) with great success.
The idea is that instead of trying to analyze a board position (which can be really, really difficult) using clever algorithms, you ask a random/simplistic algorithm to play out the rest of the game thousands upon thousands of times and see how many of those games it wins. The more it wins the better the positions.
Sounds crazy, but it actually works better than anything else.
(MCTS is usually thought of as using just one playout algorithm, with many random parameters; but that is still the same basic idea as ensembles using a bunch of different algorithms/models.)