Establishing method of dual reserve pooling nerve network model

Abstract

The invention discloses an establishing method of dual reserve pooling nerve network model, which establishes a dual reserve pooling network suitable for the decoupling confrontation performance of time sequence prediction field. The network can separate two performances having confrontation in a reserve pooling calculating model-nonlinear mapping capability and short-time memory capability. A conventional reserve pooling calculating model is based on a single reserve pooling network, like an echo state network and a liquid state machine. Although the convention reserve pooling calculating model can be applied in a dynamic system model, the calculating model needs to balance between the nonlinear mapping capability and the short-time memory capability. The invention is advantageous in thatthe hidden layer in the echo state network is arranged as two dual reserve pools, and through the main constituent analysis of a non-supervising encoder, the compressed transmission of the internal information of the reserve pool can be realized; with the dual reserve pool network, the separating and enhancing of the nonlinear mapping capability and the short-time memory capability can be realized, and good effect in the prediction of chaotic time sequence can be obtained.

Description

technical field [0001] The invention relates to the technical field of reserve pool calculation and neural network research, in particular to a method for constructing a dual reserve pool neural network model. Background technique [0002] From the perspective of dynamic system modeling, the echo state network represented by the reserve pool computing model has two important properties—non-linear mapping ability and short-term memory ability. Higher nonlinear mapping performance means that it can better fit nonlinear data, while stronger short-term memory means that the system is more inclined to reflect the law of data in the previous period, and lacks the ability to make accurate predictions for complex nonlinear data. ability. The above two abilities are mutually antagonistic in nature. When the reserve pool network shows strong nonlinearity, its memory ability will be weakened; conversely, the reserve pool network with strong memory ability will also show weak non-linea...

AI Technical Summary

Problems solved by technology

Although the existing technologies can improve the nonlinear mapping ability or short-term memory ability in the echo state network to a certain level, they have not proposed a method to jointly optimize the two confrontation performances, and cannot remove the gap between the two performances. The nature of confrontation conflicts greatly limits the ability of the current reserve pool computing network with a single reserve pool to model and predict chaotic time series

Images