Boosting Active Learning to Optimality: A Tractable Monte-Carlo, Billiard-Based Algorithm