GVA-BLS: Gaussian Vector Autoregression Broad Learning System Based on Randomly Distributed Embedding for...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Recurrent broad learning systems (RBLS) integrate broad learning systems (BLS) and recurrent neural networks (RNNs) for transient data learning.
• Traditional RBLS models inherit limitations from RNNs, including excessive weights, opaque structures, and inefficiency.

Purpose of the Study:

• To introduce a Gaussian vector autoregression BLS (GVA-BLS) model to enhance RBLS for multistep-ahead time series prediction.
• To address the shortcomings of traditional RBLS, such as excessive recurrent weights and initialization difficulties.

Main Methods:

• Developed the GVA-BLS model by integrating nonlinear vector autoregression with RBLS theory.
• Employed Gaussian kernel functions in GVA nodes for nonlinear transformation of feature node outputs.
• Mapped feature node outputs into higher dimensional spaces using GVA nodes.

Main Results:

• GVA-BLS effectively addresses issues with excessive recurrent weights and hyperparameter initialization in RBLS.
• The use of Gaussian kernels enhances GVA-BLS efficiency.
• GVA-BLS demonstrates superior multistep-ahead prediction performance compared to existing RNNs and BLS models.
• Achieved reduced computational overhead and improved intuitiveness for RBLS.

Conclusions:

• GVA-BLS offers significant improvements over traditional RBLS for time series prediction.
• The model is particularly well-suited for short-term, high-dimensional, multistep-ahead forecasting.
• GVA-BLS presents a more efficient, intuitive, and performant alternative to existing RBLS and RNN-based approaches.