Bidirectional stochastic configuration network for regression problems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Science

Background:

• Stochastic Configuration Networks (SCN) offer GPU-free training for randomized neural networks on resource-constrained devices.
• SCN's supervisory random mechanism provides stability but results in slow training times.

Purpose of the Study:

• To develop a more efficient training method for SCNs.
• To improve the training speed and hidden node quality of randomized neural networks.

Main Methods:

• Proposed a novel bidirectional SCN (BSCN) algorithm with forward and backward learning modes for adding hidden nodes.
• Forward learning uses SCN's supervisory mechanism; backward learning calculates parameters based on residual error feedback.
• Iterative application of learning modes until prediction error is acceptable or maximum hidden nodes are reached.

Main Results:

• BSCN significantly speeds up training efficiency compared to SCN.
• The bidirectional approach improves the quality of hidden nodes.
• Experiments demonstrated faster training, higher stability, and better generalization for BSCN across various benchmark and real-world problems.

Conclusions:

• BSCN offers a substantial improvement over SCN in training speed and model performance.
• The proposed semi-random learning mechanism effectively addresses the limitations of SCNs.
• BSCN is a promising approach for efficient randomized neural network training in industrial applications.