An Anthropocentric Tool for Decision Making Support

Elisabeth Le Saux, Philippe Lenca, Philippe Picouet, Jean-Pierre Barthélemy

In the complex landscape of industrial process control, the traditional approach has often been to automate human operators out of the loop to minimize error. However, researchers from ENST Bretagne propose a paradigm shift: an "anthropocentric" approach that places the expert at the heart of the machine learning process. Their tool, COMAPS, leverages cognitive modeling to assist industrial experts in maintaining and updating process control rules in real-time.

The Challenge of Industrial Process Control

Manufacturing environments—such as those producing brake pads, circuit boards, or copper foil—frequently face a conflict between "black-box" machine learning models and human intuition. While AI can process vast amounts of data, it often generates high-dimensional rules that are impossible for a human to audit or understand. Conversely, human experts possess invaluable "know-how" that is difficult to formalize and update as factory conditions evolve.

The goal of COMAPS is to bridge this gap by creating a system that captures expert knowledge without exhaustion, updates rules online as new data arrives, and respects human cognitive limits to ensure rules remain manageable.

Methodology: Bounded Rationality and the Geometric Paradigm

The COMAPS tool is grounded in two core cognitive theories:

  1. Bounded Rationality (Simon, 1979): The principle that humans make "satisficing" (good enough) decisions rather than perfect ones, constrained by short-term memory and computational capacity.
  2. Moving Basis Heuristic (MBH): A model suggesting that humans typically focus on a small subset of attributes (usually four or fewer) when making decisions.

The Geometric Paradigm

To align with these cognitive constraints, COMAPS represents the decision space as an $n$-dimensional grid. A control rule is visualized as a "cylinder" (a hypercube). If a rule depends on only two parameters in a three-dimensional space, it appears as a cylinder extending infinitely along the third, "free" dimension.

Geometric representation of a decision rule as a 'cylinder' in a multi-dimensional space.

By limiting rules to approximately four dimensions, the algorithm drastically reduces the search space, making real-time updates feasible while ensuring the resulting rules stay within human cognitive limits.

The COMAPS Algorithm

The tool operates through three distinct phases:

  • Learning: Initial rules are extracted from historical data using standard decision trees.
  • Maintenance: This is the system's core innovation. As new situations occur, the system employs seven specific heuristics (such as "generalization" or "restriction") to adjust the cylinders.
  • Conflict Solving: This phase is triggered when structural changes in the industrial process render old rules obsolete.

To prevent "knee-jerk" reactions to noisy data, COMAPS utilizes a "Wait and See" policy. It only suggests a rule modification once a sufficient number of similar cases have accumulated in a specific neighborhood of the decision space.

Key Results and Performance

The tool was validated using real-world data from industrial processes, including a dataset of 1,086 control situations updated with 556 new incoming cases. The primary metric for success was the Cylinder Factor (CF), a ratio measuring how effectively a rule covers the relevant data.

Improvement in Cylinder Factor (CF) scores for rules after the maintenance phase.

As shown in the results, the maintenance phase significantly improved rule quality. For instance, Rule 1 saw its CF score rise from 0.41 to a perfect 1.0. The system successfully integrated hundreds of new cases by applying generalizations and priority shifts, keeping the rule set concise and human-readable.

Conclusion

COMAPS demonstrates that cognitive constraints can serve as powerful computational shortcuts. By prioritizing stability over absolute perfection and keeping the "human in the loop" as a filter, the system ensures that the decision space remains logical and grounded in physical reality. Future work aims to refine priority modeling for rule intersections and conduct full-scale live industrial benchmarking.

Summary updated

BibTeX
@inproceedings{saux1999anthropocentric,
  author = {Elisabeth Le Saux and Philippe Lenca and Philippe Picouet and Jean-Pierre Barthélemy},
  title = {{An Anthropocentric Tool for Decision Making Support}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1999},
}

Loading related content...

Figure