Feedback Artificial Neural Network Research Articles

Triaxial compression behavior of sand and gravel using artificial neural networks (ANN)

Abstract The stress–strain and volume change behavior of sand and gravel under drained triaxial compression test conditions was modeled using feed-back artificial neural networks. A large experimental database obtained from published literature was used in training, testing, and prediction phases of three neural network based soil models. Issues related to the number of hidden units, magnitude of strain increment during feed-back, and over-training error are discussed. These models can accurately represent the effects of mineralogy, grain shape and size distribution, void ratio, and confining pressure. The observed behavior in terms of a non-linear stress–strain relation, compressive volume change at low stress levels, and volume expansion at high stress levels are captured well by these models.

Solving optimization problems with mean field methods

A brief review is given for the use of feed-back artificial neural networks (ANN) to obtain good approximate solutions to combinatorial optimization problems. The key element is the mean field approximation (MFT), which differs from conventional methods and “feels” its ways towards good solutions rather than fully or partly exploring different possible solutions. The methodology, which is illustrated for the graphs bisection and knapsack problems, is easily generalized to Potts systems. The latter is related to the deformable templates method, which is illustrated with the track finding problem. The mean field approximation is based on a variational principle, which also turns out to be very profitable when computing correlations in polymers.