Introduction
A fundamental challenge in artificial intelligence and engineering is the automated construction of "Goldilocks" models for physical systems—models that are sufficiently detailed for accuracy yet simple enough for computational efficiency. This paper, a classic from the 1994 AAAI proceedings, introduces an algorithm designed to bridge the gap between over-simplified and overly complex simulations by automatically selecting the simplest adequate set of model fragments for a specific query.
The Challenge of Compositional Modeling
Engineers simulating complex systems, such as power plants or spacecraft, typically utilize compositional modeling. This approach involves a library of "model fragments"—modular descriptions of physical phenomena like fluid flow, thermal resistance, or battery degradation.
The primary difficulty lies in selection. Including every possible physical phenomenon leads to a computational nightmare, while omitting a crucial detail renders the simulation results useless. This is particularly challenging for dynamic simulations, where a detail that appears irrelevant initially (e.g., battery overcharge) may become critical as the system state evolves over time.
Methodology: The select-scenario-model Algorithm
The authors propose the select-scenario-model algorithm, which leverages two primary types of relevance reasoning:
- Causal Relevance: Starting with a user query (e.g., "What is the voltage of Battery A?"), the algorithm works backward through causal influences. It identifies phenomena that directly affect the target variable and recursively identifies what influences those phenomena, building a chain of causal dependencies.
- Structural Relevance (Assumption Classes): Model fragments are organized into "assumption classes"—sets of alternative ways to model the same physical entity. For instance, a battery might be modeled as a constant voltage source or a complex, temperature-sensitive chemical cell.
The algorithm follows a "simplest-first" strategy. It begins by selecting the most basic model fragment for every relevant phenomenon. It then iteratively checks for consistency. If a simple assumption is found to ignore a variable necessary to answer the query, the algorithm "upgrades" that component to a more complex fragment. This loop continues until a stable, "simplest adequate" model is reached.
Case Study: The Solar-Battery System
To demonstrate the algorithm, the authors utilize a domain theory library and a specific scenario involving a solar array (SA1) and a rechargeable battery (BA1). The algorithm must decide which physical details to include, such as "battery damage due to overcharge" or "temperature sensitivity," based on the specific simulation goals.
Key Insights and Results
The research yields several critical insights:
- Efficiency through Simplest-First: By starting simple and adding complexity only when necessary, the algorithm avoids the exponential explosion of combinations common in "pruning" approaches.
- Causal Approximations: The framework treats simpler models as causal approximations of complex ones, facilitating efficient navigation of the model space.
- Predictive Modeling: By reasoning about potential causal influences, the algorithm selects models that remain valid throughout a dynamic simulation, even as system states change.
The algorithm provides a theoretical guarantee: it will always produce a model that is "adequate" (containing all necessary causal paths) and "simplest" (no simpler model suffices). Furthermore, it operates in polynomial time relative to the size of the resulting model, making it viable for real-world engineering tools like the Device Modeling Environment (DME).
Limitations and Future Directions
While robust, the method assumes a coherent and complete library of model fragments; it cannot rectify missing links or contradictory equations provided by human authors. Additionally, the initial version primarily handles binary modeling choices. Future work involves extending the logic to multi-valued assumption classes and refining the determination of causal directionality from system equations.
Summary updated