We report here on GENIAL (GEnetic Neural network Imple-mentation AppLication), a genetic based recurrent neural network design tool being developed for the automatic generation of trained application specific neural network architectures that can be tailored to specific constraints in actual operating systems. Strong emphasis was placed on the implementation of a genetic based design tool that could work efficiently with very small populations and obtain adequate solutions in a very reduced number of iterations when handling complex problems. GENIAL is capable of designing and training optimal (or quasi optimal) recurrent neural network architectures that can handle sequential data, minimizing the number of connections between neurons, the number of neurons, the number of processing cycles in synchronous operating mode, deciding on the optimum number of inputs per cycle, etc. It can even combine several of these constraints. It would be very simple to include further constraints and/or minimization criteria in GENIAL. Additional features allow it to work with complex neurons and even irregular combinations of different types of neurons in the same network. In order to obtain a design, GENIAL only requires a training set containing the input output pairs desired.
展开▼
