Applying recurrent neural networks to predict the training trajectory of industrial information system operators

Abstract

This paper presents the results of a series of exploratory tests of the developed recurrent neural networks (RNN) architecture xMANN (external memory memory-augmented neural network). The performance of RNN architecture xMANN was evaluated by comparing the results of the test series with the results obtained using RNN control architecture LSTM (long short-term memory) Tests were conducted with different initial conditions: composition of input data; option of educational trajectory sampling; size of embedding vector for events; size of ‘window’ in embedding layer training; number of iterations in embedding layer training; number of batches in RNN training; number of epochs in RNN training; Percentage of training examples; Fraction of trajectories used for cross-validation of RNN training epoch results. The result was a configuration of the raw data and hyperparameters of recurrent neural networks training and features of its architecture such that they enabled the prediction of the optimal learning trajectory with the maximum value of the accuracy function and the minimum value of the loss function. The xMANN architecture showed an advantage over the LSTM architecture in predicting the optimal educational trajectory accuracy for all series of experiments. On average, the loss function value for xMANN is 0.34 lower than LSTM for all series of trials, the ‘area under curve’ indicator is on average 0.07 higher than LSTM.