Networks Can Be Challenging and Time - Consuming . Experiments with Two Real - Time Applications Compared Three Approaches for Implementing

Soft computing—genetic algorithms, fuzzy sets, chaos theory, expert systems, and artificial neural networks—provides a simple way to solve complex problems. ANNs, in particular, have stimulated many theoretical studies and experiments. Researchers have proposed several architectures, but the most popular is multilayer perceptron. MLP networks implement a correspondence between input and output vectors through an assigned function with many parameters (weights). Computing these weights is an optimization problem that an analytical approach usually cannot solve. An MLP network can model the unknown function by reducing the optimization problem to a nonlinear problem of dimension equal to the number of ANN parameters. In practical applications where the training data set is large, execution times on serial machines can be extremely long. Therefore, researchers have initiated several projects to explore parallel architectures for simulating ANNs. These activities involve implementations on general-purpose parallel computers and neurocomputers (hardware dedicated to ANN simulations). Examples include ANN simulations on MAS-Par MP-1, Hypercube and Connection machines, and other supercomputers. Other groups have designed and built parallel systems based on field-programmable gate arrays, transputers, or digital signal processors (DSPs). Several companies have proposed custom-designed VLSI circuits that act like ANN accelerators—for example, CNAPS (Coprocessing Node Architecture for Parallel Systems) from Adaptive Solutions, My-Neupower from Hitachi, and Synapse-1 from Siemens. Researchers have also focused on developing faster algorithms. In the past few years, the Microcomputer Laboratory at the University of Pavia has begun researching hardware solutions for heavy computing problems. This activity involves ANNs used in typical real-time industrial applications, such as signal filtering or compression, recognition, and control. We have implemented these ANNs on either workstations or architectures that integrate existing microcomputers and microprocessors. Here, we compare the performance of three such solutions for two real-time industrial applications and identify the most promising way to train and test an MLP neural network. Giovanni Danese Francesco Leporati