Information-Driven Rule Reduction in Belief Rule Bases for Complex System Modeling.

Area of Science:

• Engineering Systems Analysis
• Machine Learning
• Uncertainty Quantification

Background:

• Managing uncertainty and optimizing information processing are crucial for reliable state prediction in complex engineering systems.
• Belief Rule Base (BRB) integrates expert knowledge with uncertain information but faces combinatorial complexity issues.
• Conventional BRB simplification can lead to information loss and prediction biases, compromising reliability.

Purpose of the Study:

• To propose an adaptive belief rule base framework (BRB-ARR) that balances system complexity and modeling accuracy.
• To mitigate prediction biases and information loss caused by conventional BRB structure simplification.
• To enhance computational efficiency and preserve the interpretability of the inference architecture.

Main Methods:

• Developed an information-driven rule screening mechanism using optimized Mean Square Error (MSE) fluctuations to dynamically prune redundant rules.
• Employed a low-dimensional optimization process to readjust the parameter vector for improved computational efficiency.
• Introduced a posterior calibration module to compensate for systematic biases resulting from dimensionality reduction.

Main Results:

• In petroleum pipeline networks, the rule base scale decreased by over 60% (56 to ~20 rules) and parameter dimensionality reduced from 338 to 122.
• The mean squared error (MSE) for petroleum pipelines improved from 0.5291 to 0.3619.
• For liquid propellant launch vehicles, prediction accuracy reached 98.57% with an MSE of 0.00029, reducing rule scale from 441 to 109.

Conclusions:

• The BRB-ARR model effectively balances structural compactness with high-precision prediction.
• The framework offers a novel approach to uncertainty modeling in intelligent systems.
• Experimental results validate the framework's effectiveness in complex engineering applications.