Dealing with Uncertainty in Fuzzy Inductive Reasoning Methodology

Francisco Mugica, Àngela Nebot, Pilar GómezUAI 2003

TL;DR

This paper introduces a "mixed" prediction scheme for Fuzzy Inductive Reasoning (FIR) that combines the accuracy of pattern-based rules with the efficiency of compact Sugeno fuzzy rules. This approach aims to strike a better balance between prediction performance and model size, particularly in systems with varying levels of uncertainty.

What Problem Does This Paper Address?

Fuzzy Inductive Reasoning (FIR) is a data-driven methodology for modeling systems where structural knowledge is limited. It works by first identifying patterns in data and then using these patterns to predict future system behavior. However, FIR has a drawback: the "pattern rule base" it generates can become excessively large, especially with abundant data.

To address this, the CARFIR (Construction of Automatic Fuzzy Rules) methodology was developed. CARFIR compacts the large pattern rule base into a smaller set of Sugeno fuzzy rules, significantly reducing model size. While this compaction improves efficiency, it can lead to a loss of prediction accuracy, particularly in areas of the data space where the original patterns exhibit high uncertainty. This paper tackles the challenge of how to best leverage both the accuracy of pattern rules and the efficiency of Sugeno rules when dealing with such uncertainty.

Methods

The paper proposes a mixed pattern/fuzzy rules strategy to handle uncertainty. This strategy builds upon the CARFIR methodology, which first generates a pattern rule base from data and then automatically extracts Sugeno fuzzy rules from it.

The core of the proposed method involves:

  • Error Modeling: Identifying areas of high uncertainty by comparing the predictions from the original pattern rule base (FIR) and the compacted Sugeno fuzzy rule base (FIS). Three error models (G1, G2, G3) are used to quantify these differences.
  • Selective Pattern Rule Preservation: Based on the error models, a subset of the original pattern rules is selectively kept. These are the rules located in regions identified as having high uncertainty.
  • Mixed Prediction Scheme: For regions with high uncertainty, a combination of the preserved pattern rules and the Sugeno fuzzy rules is used for prediction. For regions with low uncertainty, only the Sugeno fuzzy rules are employed.
  • Weighting Mechanism: The contribution of pattern rules versus Sugeno rules in the mixed scheme is determined by a weighting function (fmix) that depends on the normalized Euclidean distance between the current system state and the closest preserved pattern rule. This function allows for a smooth transition between relying more on pattern rules in uncertain areas and more on fuzzy rules in certain areas.

Datasets

The proposed methodology is evaluated on a biomedical application: modeling the human Central Nervous System (CNS) control of the cardiovascular system. The input data used for training and testing consists of recorded signals from simulations of a differential equation model of the CNS, specifically tuned to represent a patient with coronary arterial obstruction. The data was collected with a sampling rate of 0.12 seconds. The paper mentions using 7277 samples for training and six test data sets, each with approximately 600 data points, for validation.

Key Insights and Ideas

  • Uncertainty as a Guide: The paper highlights that uncertainty in the original pattern rules is a key indicator of where compacting to Sugeno rules might lead to significant accuracy loss.
  • Targeted Accuracy Recovery: Instead of abandoning the efficiency of Sugeno rules entirely, the mixed scheme intelligently reintroduces pattern rules only in the most uncertain regions, aiming for a localized accuracy boost.
  • Trade-off Management: The mixed scheme provides a tunable mechanism to balance prediction accuracy against model size, allowing users to prioritize either speed or precision based on their needs.
  • Error Models as Differentiators: The three error models (G1, G2, G3) provide different perspectives on the discrepancy between pattern and fuzzy rule predictions, with G2 and G3 showing particular promise in guiding the mixed scheme.

Key Results

  • Significant Rule Reduction: The CARFIR methodology achieved a substantial reduction in rules, from 7275 pattern rules to 81 Sugeno fuzzy rules per controller.
  • Accuracy Loss with Pure Sugeno: The pure Sugeno prediction scheme resulted in higher Mean Squared Errors (MSE) compared to the pattern prediction scheme. For example, the average MSE for the coronary resistance controller was 2.91% with Sugeno rules, versus 1.63% with pattern rules.
  • Mixed Scheme Performance: The mixed scheme demonstrated an ability to recover prediction accuracy as the percentage of preserved pattern rules increased. For the PR controller, preserving 20% of pattern rules (using the G2 error model) recovered 80% of the accuracy lost by using only Sugeno rules.
  • Optimal Compromise: For the specific application, preserving around 30% of pattern rules (or achieving a 70% rule base reduction) was identified as an optimum value balancing accuracy and computational cost.
  • Comparison to Other Methods: The CARFIR methodology (both Sugeno and mixed schemes) generally outperformed NARMAX, TDNN, and RNN models in terms of prediction accuracy for the CNS control problem.

Limitations and Future Work

  • Generalizability of Mixed Scheme: The authors acknowledge that defining a universal decision procedure for selecting the percentage of pattern rules to keep in the mixed scheme is difficult, as it highly depends on the specific application and user interests.
  • Future Applications: The authors suggest using CARFIR in applications with even larger datasets to explore its full potential.
  • Theoretical Study: A theoretical study on rule base reduction is planned for future work.

Summary updated

BibTeX
@inproceedings{mugica2003dealing,
  author = {Francisco Mugica and Àngela Nebot and Pilar Gómez},
  title = {{Dealing with Uncertainty in Fuzzy Inductive Reasoning Methodology}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2003},
}
Figure