A CAD approach to the design of a neural network simulator

This paper shortly describes a preliminary version of a new discrete-time network software simulator implemented using the signal-flow-graph (SFG) approach. The software is able to do a very user-friendly schematic entry of the network and the cost function, and can implement any discrete-time network composed by delays, linear and nonlinear fired or adaptable components (also user-defined). Structures such as usual linear filters (FIR, IIR) and static multilayer, time delay or recurrent neural networks of arbitrary topologies can be easily adapted with respect to any user-selected parameters, minimizing any objective function that can be described by a SFG (online, batch, windowed MSE, etc.) using the BC method.