Violation-guided learning for constrained formulations in neural-network time-series predictions
Benjamin Wan-Sang Wah, Minglun Qian · International Joint Conference on Artificial Intelligence · 2001
Time-series predictions by artificial neural networks (ANNs) are traditionally formulated as unconstrained optimization problems. As an unconstrained formulation provides little guidance on search directions when a search gets stuck in a poor local minimum, we have proposed recently to use a constrained formulation in order to use constraint violations to provide additional guidance. In this paper, we formulate ANN learning with cross-validations for time-series predictions as a non-differentiable nonlinear constrained optimization problem. Based on our theory of Lagrange multipliers for discrete constrained optimization, we propose an efficient learning algorithm, called violation guided back-propagation (VGBP), that computes an approximate gradient using back-propagation (BP), that introduces annealing to avoid blind acceptance of trial points, and that applies a relax-and-tighten (R&T) strategy to achieve faster convergence. Extensive experimental results on well-known benchmarks, when compared to previous work, show one to two orders-of-magnitude improvement in prediction quality, while using less weights.