Hybrid combination genetic algorithm and controlled gradient method to train a neural network

Abstract

Multivariate predictive analysis is a widely used tool in the petroleum industry in situations in which the deterministic nature of the relationship between a variable that requires prediction and a variable that is used for the purposes of such prediction is unknown or very complex. For example, to perform a sweet-spot analysis, it is necessary to predict potential oil and gas production rates on a map, using various geologic and geophysical attribute maps (porosity, density, seismic attributes, gravity, magnetic, etc.) and the initial oil and gas production rates of several control or training wells located in the area of interest. We have developed a new technology that allows for building a stable nonlinear predictive operator by using the combination of a neural network, a genetic algorithm, and a controlled gradient method. The main idea behind the proposed technology is to combine stochastic and deterministic approaches during the construction of the predictive operator at the training stage. The proposed technology avoids many disadvantages of the genetic algorithm and gradients methods, such as a high level of dependency on the initial values; the phenomenon of over-fitting (overtraining), which results in creation of an operator with unstable predictability; and a low speed of decreasing error during iteration, and, as a result, a low level of prediction quality. However, the above-mentioned combination uses the advantages of both methods and allows for finding a solution significantly closer to a global minimum for the objective function, compared to simple gradient methods, such as back propagation. The combination of these methods together with Tikhonov regularization allows for building stable predictions in spatial or/and time coordinates.