Self-Explanatory Simulations: Scaling Up to Large Models

Kenneth D. Forbus, Brian FalkenhainerAAAI 1992

Abstract

This paper presents SIMGEN.MK2, a system for generating self-explanatory simulations that combine numerical precision with qualitative reasoning. By transitioning from global state envisionment to local reasoning and bounded environments, SIMGEN.MK2 overcomes the combinatorial explosion that previously limited qualitative modeling. Experimental results demonstrate that the system can handle models with an estimated trillion states, which were previously computationally intractable.

Introduction

Self-explanatory simulations represent a hybrid approach to modeling, combining the numerical precision of traditional simulators with the explanatory power of qualitative physics. In the early 1990s, the field of qualitative reasoning (QR) faced a significant "scaling wall." The prevailing methodology, exemplified by SIMGEN.MK1, relied on total envisionment—the process of pre-calculating every possible global state a system could inhabit. While feasible for simple systems, the state space grows exponentially with complexity, making global state reification computationally prohibitive for real-world engineering applications.

Methodology: From Global to Local Reasoning

SIMGEN.MK2 shifts the paradigm from global state mapping to efficient local reasoning. Key innovations include:

  1. Local State Analysis: Instead of a global map, the system identifies "model fragments" (e.g., fluid flows, heat paths) and their local conditions.
  2. Assumption-based Truth Maintenance System (ATMS): To manage local dependencies without exponential growth, the authors implemented bounded environments ($E_{max}$), limiting the number of concurrent assumptions tracked.
  3. Unified Code Generation: MK2 generates a single "evolver" for numerical integration and a transition procedure for qualitative changes, rather than separate code for every global state.
  4. Concise Histories: Explanations are generated on-the-fly from a history of boolean variables, rather than being retrieved from a pre-computed state map.

Evaluation and Results

The system was evaluated against small benchmarks and a complex "3x3 grid" scaling test. The 3x3 grid, featuring nine containers and twelve fluid paths, would require an estimated $10^{12}$ states for total envisionment. SIMGEN.MK2 successfully analyzed this model in approximately 4.5 hours, with code generation taking less than 100 seconds.

On smaller benchmarks, MK2 demonstrated significant performance gains over the envisioning-based MK1 approach.

Performance comparison (in seconds) of SIMGEN.MK2 components versus traditional envisioning (MK1) for small benchmarks.

As shown in the table above, the time required for qualitative analysis in MK2 is consistently lower than the time required for total envisionment in the previous generation. For the "Two containers" model, the qualitative analysis phase was completed in 19.4 seconds, compared to 40.2 seconds for the old envisioning method. Even when adding the code generation time (3.4 seconds), the MK2 system remains significantly more efficient.

Discussion and Conclusion

By abandoning the requirement for total envisionment, SIMGEN.MK2 enables the creation of self-explanatory simulators for systems previously considered untouchable. While this approach sacrifices "cheap counterfactuals"—the ability to instantly query alternative histories—it gains the ability to scale to real-world industrial complexity. Future work focuses on optimizing qualitative analysis routines and integrating the system into safety-critical training environments where explainability is paramount.

Summary updated

BibTeX
@inproceedings{forbus1992self,
  author = {Kenneth D. Forbus and Brian Falkenhainer},
  title = {{Self-Explanatory Simulations: Scaling Up to Large Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure