bayhess 0.1.5

Bayesian Hessian Approximation for Stochastic Optimization

The BayHess package uses noisy curvature pairs (noisy gradient differences computed at different points) to compute Hessian approximations. These Hessian approximations can be used to accelerate the convergence in stochastic optimization in a quasi-Newton fashion. To find a Hessian approximation, a posterior distribution of the Hessian is built. The prior distribution is based on the Frobenius norm with determinant constraints to impose extreme eigenvalues constraints and the likelihood distribution is built from the secant equations given the observed curvature pairs. To find the maximizer of the log posterior, the BayHess package uses the Newton-CG method with a homotopy approach to deal with the logarithmic barrier determinant constraints.

For a detailed description of the method, convergence analysis and numerical results, check our manuscript named “Approximating Hessian matrices using Bayesian inference: a new approach for quasi-Newton methods in stochastic optimization”. This package can be used with the MICE estimator.