Применение нейронных сетей прямого распространения для формирования оценок параметров по методу максимального правдоподобия

Наукоемкие технологии. 2015. № 8. С. 42–47.

Филимонова Д. В., Shevgunov T., Ефимов Е. Н.

This paper introduces an approach for implementing maximum likelihood parameter estimators using feedforward artificial neural network of multilayer perceptron architecture. A theoretical foundation of the proposed approach is presented in the assumption that the model of observation is known as well as the values of its vector of parameters. For a practical example the implementation of direction of arrival estimator for the active ring antenna array is shown. In order to estimate a performance and accuracy of the proposed approach, the results of numerical calculation are presented, compared to the algorithm based on optimal numerical solution and referenced to Cramer–Rao lower bound. The results also indicate that there is no significant dependency of the accuracy of estimation on actual parameter value. Moreover, the calculations take significantly less time, although some of it is spent on the initial training of the neural network.

Keywords: artificial neural networks, искусственные нейронные сети, maximum likelihood estimator, метод максимального правдоподобия, circular antenna arrays, угломерный метод, кольцевые антенные решетки, parametric estimation, DOA, параметрическое оценивание