Introduction
In the early 1980s, the rise of Artificial Intelligence enabled the creation of highly detailed simulations, such as the ROSS air warfare simulator. These models could track every individual radar detection, fighter assignment, and tactical maneuver. However, this high fidelity introduced two significant problems: extreme computational costs as simulation scale increased, and "cognitive overload" for human users who struggled to find the "big picture" amidst an agonizingly detailed trace of events.
To address this, researchers at the Rand Corporation developed the Abstracted Simulation Model (ASM). The goal was to create "quick and dirty" approximate simulations that maintain the essential logic of a detailed model while running at a fraction of the cost and providing more intelligible outputs.
The Abstracted Simulation Model (ASM)
The ASM operates through a two-phase process: a learning phase where it observes a detailed model, and an event generation phase where it runs independently.
1. The Learning Phase
Rather than requiring manual coding of abstract rules, the ASM "watches" the Detailed Simulation Model (DSM) to identify patterns. It groups events into "interaction sets" based on the roles involved (e.g., Radars and Bombers) and monitors transition probabilities between these events.
Hierarchical Chunking
The core innovation of the ASM is "chunking." When the system identifies a sequence of events that occurs frequently, it collapses them into a single high-level macro-event. For instance, a sequence of "detect, identify, and track" might be abstracted into a single "intercept" event.
As shown in the hierarchy above, this allows the simulation to represent complex processes at multiple levels of granularity, depending on the user's needs.
2. The Event Generation Phase
Once the learning phase is complete, the ASM functions as a probabilistic scenario network (similar to a Markov model). It moves from one state to another based on learned probabilities rather than re-calculating heavy heuristic rules.
These scenario networks can also trigger one another. For example, a specific interaction in a bomber/radar subnetwork might act as a constraint that "activates" a separate radar/command-center subnetwork, as illustrated in the figure above.
Key Insights and Results
The ASM approach offers several advantages over traditional high-fidelity simulations:
- Cognitive Economy: By drawing a parallel to human "scripts" (e.g., remembering the concept of "going to a restaurant" rather than every individual muscle movement), the ASM provides a summary that is naturally easier for humans to process.
- Computational Efficiency: Because the ASM follows a probabilistic path through a pre-learned network, it bypasses the heavy computational load of the DSM's heuristic rule-checking.
- Information Filtering: Users can choose their desired level of abstraction, effectively using the ASM as an intelligent filter to track only specific roles or high-level outcomes.
Limitations and Future Directions
While promising, the current ASM framework has limitations. It relies on human intuition to define the initial "elementary events" and cannot yet "discover" entirely new concepts outside of those present in the training data. Furthermore, the model is less flexible than a DSM; changing physical parameters like speed or range currently requires retraining the network.
Despite these hurdles, the ASM provides a vital bridge between high-speed computation and human-scale decision-making, transforming data-heavy traces into actionable insights.
Summary updated