An Anthropocentric Tool for Decision Making Support

TL;DR

This paper introduces COMAPS, a novel decision support tool for industrial process control that prioritizes the human expert's knowledge and cognitive limitations. By modeling expert decision-making processes, COMAPS aims to create more flexible and adaptable machine learning systems that integrate seamlessly with human operators, rather than replacing them.

What Problem Does This Paper Address?

Traditionally, industrial process control systems have viewed human operators as a primary source of errors. However, there's a growing recognition that the unique knowledge and experience of human experts are invaluable resources. The challenge lies in effectively capturing and leveraging this "know-how" within automated decision-making systems, especially when dealing with dynamic industrial processes that require on-the-fly adaptation. Existing machine learning and total quality management tools often fall short because they don't fully account for the flexibility and reusability of human expertise, nor do they easily allow for human intervention. This paper addresses the need for tools that can integrate human cognitive principles to support, rather than supplant, expert decision-making in real-time industrial control.

Methods

The COMAPS tool is built upon two core cognitive models: Bounded Rationality (Simon, 1979) and the Moving Basis Heuristic (MBH) (Barthelemy & Mullet, 1992). These models suggest that experts operate with a limited set of stable strategies, focusing on a small subset of relevant attributes, and aiming for satisfactory rather than optimal solutions.

COMAPS models the decision-making process through a geometric paradigm in an n-dimensional space, where each point represents a "parameter setting." Expert strategies are represented as cylinders in this space, defined by a conjunction of attribute aspects. The system operates in three phases mirroring the process lifecycle:

  1. Learning Phase: Extracts initial rules (cylinders) from historical data using adapted machine learning techniques.
  2. Maintenance Phase: The core of the paper, this phase involves on-line updating of the cylinder "paving" (the set of rules) as new control situations arise. This phase is designed to be cognitively constrained, aiming for a small number of dimensions and cylinders.
  3. Conflict Solving Phase: Handles situations where no acceptable modification can be made to update the paving, often triggered by significant technological changes.

A key innovation is the handling of NP-complete problems (like finding maximal cylinders or checking paving consistency) through heuristics and a priority system between overlapping cylinders. This priority system is crucial for managing potential intersections and ensuring a coherent decision-making process.

Datasets

The paper mentions testing the algorithm on three real industrial processes (copper foil production, printed circuit board manufacturing, and brake pad production) within the COMAPS project. However, due to the confidential nature of real data, the initial tests were conducted using a mock-up with coded data. The paper details experiments using a training set and an incoming situations set derived from this coded data, with a history of 1086 control situations and 556 new situations used for testing the maintenance phase.

Key Insights and Ideas

  • Anthropocentric Approach: The fundamental idea is to put the human expert back at the center of the decision process, leveraging their knowledge as a resource rather than an error source.
  • Cognitive Constraints are Beneficial: Instead of trying to overcome human cognitive limitations (like short-term memory capacity), the system is designed to work within them, making it more natural for experts to use and understand.
  • Cylinders as Expert Rules: Representing expert rules as geometric cylinders in a multi-dimensional parameter space provides an intuitive and computationally manageable way to model decision strategies.
  • Heuristics for NP-Completeness: The paper highlights the necessity of using heuristics to bypass computationally intractable problems (NP-complete) in real-time decision support, making the system practical.
  • On-line Adaptation: The maintenance phase's ability to continuously update the decision rules based on new incoming data is critical for adapting to evolving industrial processes.
  • Priority for Overlapping Rules: The introduction of a priority mechanism between potentially overlapping cylinders is a clever way to resolve ambiguities and maintain a consistent decision output.

Key Results

The paper presents results from testing the maintenance phase on coded data. Starting with 6 institutional rules and a history of 1086 control situations, the ruleset was updated based on 556 new control situations. The quality of the rules, measured by a "CF" (Covering Factor) metric, showed significant improvement after the maintenance phase. For example, Rule 1 improved its CF from 0.41 to 0.84, and Rule 3 from 0.56 to 0.87. These improvements were validated by the expert. The paper also notes that the modifications applied were categorized according to the functionalities used (e.g., generalization, creation, restriction).

Limitations and Future Work

  • Internal Model of Priority: The authors acknowledge that the current internal model for cylinder priority is a simple boolean matrix and lacks information for complex intersections of more than two cylinders. They plan to improve this.
  • Real-time Benchmarking: While promising, the current tests were on coded data. Future work involves testing on a prototype installed in a real industrial setting for more robust validation.
  • Comparison with Classical ML: Comparisons with established decision tree learning tools like ID3 and C4.5 are underway.
  • Human-AI Interaction: The paper implicitly suggests that further work could explore more nuanced ways for experts to interact with and refine the system's decision-making process.

Summary updated

BibTeX
@inproceedings{1999anthropocentric,
  title = {{An Anthropocentric Tool for Decision Making Support}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1999},
}
Figure