 |
Center for Nonlinear Studies |
In this talk I explore the relationship between Machine-Learning (ML), Monte Carlo Optimization of a parametrized integral (MCO), and `blackbox' optimization (BO - the kind of optimization one typically addresses with GA's or simulated annealing). First, I show how to transform any BO problem into an MCO problem. Second, I show that MCO is formally identical to ML. These two results provide a way to apply many powerful ML techniques to BO. In particular, I show how to exploit cross-validation and bagging for BO. I end by showing how using any (?) type of Monte Carlo to estimate an integral can be improved by using the ML technique of stacking.
Host: CNLS