Deep Reservoir Computing Model And Its Application In Blind Equalization Of Communication Systems

Reservoir computing is an effective method for training recursive neural networks(RNNs),with classical RNN models based on reservoir computing methods including Echo State Networks(ESN)and Liquid State Machines(LSM).Deep reservoir computing models have superior regression performance compared to basic reservoir computing models,through simple stacking and ensemble techniques,and demonstrate broader applicability in problems such as chaotic time-series signal prediction and nonlinear system modeling.In this paper,a deep Echo State Network model based on Conditional Gaussian Restricted Boltzmann Machine(CGRBM)was proposed on the basis of the classical ESN model,and applied to the blind equalization problem in communication systems.The main contributions and research results of the paper are as follows:(1)We propose a deep Reservoir Echo State Network model,C-R-DESN,based on Conditional Gaussian Restricted Boltzmann Machines(CGRBMs).This model utilizes CGRBMs to extract latent features from input data.Additionally,it incorporates the Random Forest algorithm from ensemble learning to prune redundant features within the set of latent features.This process generates multiple subsets of significantly diverse features.Finally,the ESN model replaces the decision trees in Random Forest as the base learners in the ensemble.Simulation results demonstrate the excellent performance of the proposed algorithm in both time series prediction and system modeling tasks.(2)A C-R-DESN model based on a heuristic search algorithm is proposed,which further improves the performance of the C-R-DESN model by using the Sparrow search algorithm(SSA)to heuristically optimize the hyperparameters in the C-R-DESN.Experimental results demonstrate that the proposed model exhibits superior regression performance on the Mackey-Glass17,Lorenz,and NARMA datasets.(3)To further extend the adaptability of ESN-based blind equalization algorithms in strongly nonlinear channels,a batch blind equalization algorithm called C-R-DESN based on Nonlinear Prediction Error Filter is proposed.The algorithm combines the C-R-DESN model with the prediction principle blind equalization method to achieve blind equalization of high-order QAM signals under a strongly nonlinear channel generated by a fifth-order Volterra stage.