Recurrent neural network optimisation based on linearly constrained numerical methods

Time-series data analysis has grown even more crucial in many sectors as information technology and big data expand rapidly. This work proposes a recurrent neural network (RNN) optimisation model based on the linear constraint numerical method, namely, LSTM-LP optimiser, which combines the powerful time-series modelling capability of long short-term memory (LSTM) and the optimisation characteristics of linear programming (LP) optimisation features, and so effectively improves the training efficiency and stability of the model in resource-constrained environments. This helps to efficiently capture the temporal dependencies in time-series data and solve the noise and missing problems in the data. On two datasets, experimental results show the LSTM-LP optimiser beats the conventional model in several performance criteria. Future studies will investigate more effective optimisation techniques, increase the generalisation capacity of the model, and simplify the hyperparameter tweaking process to thus further promote the model in practical uses.