Using Machine Learning to Improve Monte Carlo and Blackbox Optimization
- David Wolpert (Santa Fe Institute and Los Alamos National Laboratory, USA)
Abstract
In this talk I explore the relationship between Machine-Learning (ML), Monte Carlo Integral Estimation (MC), and 'blackbox' optimization (BO - the kind of optimization one typically addresses with genetic algorithms or simulated annealing).
To begin, I show how to to apply ML to automatically set annealing temperatures and other hyperparameters of *any* stochastic BO algorithm. Then I show how to extend this, first to transform any BO problem into a particular MC problem, and then to show that this MC problem is formally identical to the problem of how to do supervised learning. This extension allows us to apply *all* of the powerful techniques that have created for supervised learning to solve BO problems. I demonstrate the power of this in experiments (movies!).
I end by showing how to improve the convergence of any of a broad set of MC algorithms, by using the ML technique of stacking to learn control variates.
