John Regehr writes this post about a compiler speed optimization technique called “Stochastic Superopimization“. “Stochastic Superopimization” systematically searches for algorithmic improvement in code using machine learning algorithms similar to multi-armed bandit strategies. It appears to be related to “Programming by Optimization“. “Stochastic Superopimization” is more like a very good optimization flag on a compiler. “Programming by Optimization” is constructing the program in such a fashion that design options are exposed and easily manipulable by an optimization program trying to maximize some performance metric. The “Programming by Optimization” community seems to mostly use BOA, the Bayesian optimization algorithm (see [1], [2]). I am hoping to read and write more about both of these ideas later.