Output details

Non-Smooth Optimization Based on Resilient Backpropagation

Search for Unconstrained and Simply Bounded Problems.

This paper presents a novel use of backpropagation for determining search directions for non-smooth functions minimisation. Costly line searches and trust regions subproblems, used by subgradient and bundle methods, are avoided. This done by requiring just one function and one gradient evaluation at each iteration along with a simple comparison of gradient sign values. The results obtained are compared with state-of-the-art bundle methods and it is shown that the proposed algorithm has better convergence behaviour. Furthermore, this paper reports results for problems of many thousands dimensions, which does not appear elsewhere in the literature of non-smooth optimisation.