Abstract
Automated modeling of complex physical systems requires balancing accuracy with computational efficiency. This paper presents TRIPEL, a system that addresses the "scope" problem by automatically selecting the appropriate time scale and system boundary for a given prediction question. By filtering influences based on temporal relevance and causal connectivity, TRIPEL generates simplified, coherent models from large domain knowledge bases. Evaluation on plant physiology problems demonstrates that TRIPEL significantly reduces model complexity while maintaining predictive adequacy.
Introduction
Predicting the behavior of physical systems—such as a plant's response to changing soil moisture—is challenging because these systems are influenced by processes operating across vastly different time scales, from cellular osmosis (seconds) to seasonal climate shifts (months). Traditional modeling requires human experts to manually define the "system boundary," deciding which variables to include and which to ignore. If a model is too simple, it lacks accuracy; if too complex, it becomes computationally prohibitive and difficult to interpret. This research introduces TRIPEL, a program designed to automate these choices, ensuring models are "just right" for specific prediction tasks.
Methodology: The TRIPEL System
TRIPEL is built on the principles of compositional modeling and employs a multi-step process to derive minimal adequate models.
Influence Graphs and Time Scale Selection
The system represents domain knowledge as an influence graph, where nodes represent variables and edges represent causal relations. A core innovation of TRIPEL is its use of time as a filter. It distinguishes between differential influences (causing change over time) and functional influences (instantaneous).
To find the correct scope, TRIPEL starts at the fastest time scale and checks for a "differential influence path" from the question's input to the target variable. If no such path exists, it incrementally considers slower time scales until a valid causal chain is identified.
System Boundary and Model Generation
Once a time scale is selected, TRIPEL prunes influences that are too slow to be significant. It then performs a connectivity analysis using the Floyd-Warshall algorithm to isolate relevant variables. Finally, a best-first search identifies the simplest model that satisfies logical "coherence constraints," ensuring that the selected model fragments work together consistently.
Evaluation and Results
The system was evaluated using a plant physiology knowledge base containing 77 variables and 155 candidate influences, covering processes like water regulation and carbon dioxide exchange.
Model Simplification
TRIPEL demonstrated remarkable efficiency in reducing model complexity. For seven test questions, the system reduced the average model size from 77 variables to just 11 variables and 14 influences. This pruning allows for much faster simulation and easier human interpretation of the resulting model.
Performance and Validation
The models were generated in less than 15 seconds. In the most efficient case, connectivity analysis reduced the search space from 77 variables to just 4. Domain experts validated the resulting system boundaries and time scales, confirming they were sufficient and reasonable for the questions posed.
Discussion and Future Work
TRIPEL's primary insight is that time scales provide a natural hierarchy for physical systems, allowing for the isolation of manageable subsystems. However, the current implementation assumes a single time scale is sufficient for any given question. Future work will explore modeling interactions across multiple simultaneous time scales and scaling the system to larger knowledge bases, such as the Botany Knowledge Base with over 200 processes.
Summary updated