The Challenge of Real-Time Process Control for Production Systems

Franz Barachini, Norbert TheuretzbacherAAAI 1988

Abstract

Production systems based on the RETE algorithm have historically struggled with the stringent timing requirements of industrial process control. This paper presents PAMELA (PAttern Matching Expert system LAnguage), an expert system language designed to achieve real-time performance. By optimizing network traversal, reducing node redundancy, and introducing a novel interrupt-handling architecture, PAMELA significantly outperforms industry-standard inference engines like OPS83.

Introduction

In the late 1980s, the application of expert systems to real-time process control was limited by two factors: computational overhead and the lack of asynchronous capabilities. The standard "recognize-act" cycle of production systems is inherently serial and rigid, making it difficult to respond to external events without compromising data integrity. PAMELA was developed to overcome these barriers, providing a bridge between declarative rule-based logic and the millisecond-sensitive demands of industrial environments.

Architectural Optimizations in PAMELA

PAMELA introduces several architectural refinements to the classic RETE algorithm to enhance speed and flexibility:

  • Stack Optimization: By representing nodes as procedures and utilizing the hardware stack through recursion, PAMELA eliminates the overhead of interpreted token-stacks.
  • Backlink Pointers: The use of backlink pointers and counters allows for rapid REMOVE and CHANGE operations, bypassing the need for exhaustive network re-checks.
  • Node Reduction: A recursive optimization algorithm sorts and compresses the RETE network. This approach can reduce the total number of nodes by over 40%, streamlining the matching process.
  • The Demon Concept and Interrupt Handling: Unlike standard rules, "demons" in PAMELA fire immediately upon condition satisfaction. To handle external interrupts safely, the system utilizes atomic, queued actions and synchronization slots.
The PAMELA architecture integrates demons and asynchronous interrupt queues into the inference cycle.

As illustrated in the architecture diagram, memory modifications (Q_MAKE, Q_CHANGE, Q_REMOVE) are treated as atomic operations, ensuring that external events do not disrupt the internal logic of an active rule.

Evaluation and Results

The system was benchmarked against several classic expert system problems, including MAB-NASA (planning), COMBI (combinatorial optimization), and RUBIK (Rubik’s Cube solver). These tests compared PAMELA’s various optimization modes against OPS83, the leading inference engine of the era.

Performance comparison of PAMELA variants against OPS83 across several standard benchmarks.

The experimental results highlight the efficiency of the PAMELA engine. On the MAB-NASA benchmark, PAMELA achieved speeds twice as fast as OPS83. The implementation of two-input node sharing provided a 22% performance improvement in specific cases, although the authors noted that the benefits of sharing are highly dependent on the nature of the data and the cross-product of tokens in memory. Overall, PAMELA demonstrated superior scalability as the size of the working memory increased.

Conclusion

PAMELA successfully addresses the performance and asynchrony challenges of real-time expert systems. Through hardware-level optimizations and a robust interrupt-handling framework, it makes declarative production systems viable for complex process control tasks. As uniprocessor performance reaches its limits, the authors suggest that future advancements will likely involve parallel architectures.

Summary updated

BibTeX
@inproceedings{barachini1988challenge,
  author = {Franz Barachini and Norbert Theuretzbacher},
  title = {{The Challenge of Real-Time Process Control for Production Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure