Parametric optimization method for wide neural networks using genetic algorithms

In many complex technical problems, the number of parameters available for analysis is in the thousands. At the same time, depending on the specifics of the problem being solved, usually the number of key parameters, that is, parameters that have a significant impact on the processes associated with the target problem, does not exceed several tens. However, determining a subset of these parameters from those available in itself is a difficult task, which in most cases is solved with the assistance of experts in the relevant subject area. This paper proposes a parametric optimization method that can be used for wide neural networks, that is, neural networks with a large number of neurons in a layer. This method uses evolutionary optimization methods, namely, genetic algorithms, together with the method of invariant data representation in wide neural networks, using the algebra of hyperdimensional binary vectors, due to which, when the number of parameters of a neural network model changes during optimization, its topology does not change. At the same time, the more parameters are included in the model, the less accurately their values are transmitted, thus, in the course of optimization, a balance is achieved between the composition, number and accuracy of the parameters of the target problem. The proposed method does not require the participation of an expert corresponding to the subject area, allowing the process of parametric optimization to be fully automated.