Plan Recognition for Airborne Tactical Decision Making

Jerome Azarewicz, Glenn Fala, Ralph Fink, Christof HeitheckerAAAI 1986

Introduction

In the high-stakes environment of mid-80s aerial combat, pilots faced an overwhelming "fog of war." Sophisticated sensors provided high data rates, but the resulting uncertainty and the need for split-second responses created a significant cognitive bottleneck. This 1986 paper introduces a Plan Recognition Model (PRM) designed to assist pilots by hypothesizing enemy goals and predicting future maneuvers, even when faced with noisy or deceptive sensor data.

Unlike contemporary AI research that focused on cooperative agents in static environments, the PRM addresses the adversarial nature of tactical combat. It bridges the gap between theoretical plan recognition and the "messy" reality of multiple adversarial agents and elusive, time-critical goals.

System Architecture

The PRM is built upon three core cognitive components, adapting John Anderson’s "Architecture of Cognition" into a blackboard-based framework.

The three major cognitive components of the Plan Recognition Model (PRM): Long Term Memory, Procedural Memory, and Short Term Memory.
  1. Long Term Memory (LTM): This serves as a declarative library of missions, plans, and tactics. It utilizes a hierarchical structure to represent the legal sequences of maneuvers an enemy might perform.
  2. Procedural Memory (PM): Acting as the system's engine, PM employs an informed best-first search. It uses "event matchers" to compare real-time sensor data against the constraints defined in the LTM.
  3. Short Term Memory (STM): Implemented via a blackboard architecture, the STM is a dynamic workspace. Different software modules (experts) post observations, hypotheses, and threat estimates here, allowing for the integration of bottom-up sensor data with top-down tactical expectations.

Plan Representation and Matching

The LTM organizes tactical knowledge into a hierarchy of Plans, Events, Parameters, and Constraints. This structure allows the system to rapidly prune the search space by discarding plans that violate established constraints (e.g., a specific aircraft type exceeding its known performance envelope).

The hierarchical structure of the Plan Recognition Model, showing the relationship between plans, events, parameters, and their associated constraints.

A critical feature of the PRM is the calculation of a Degree of Match (DOM). Observed behaviors are categorized as Exact, Partial, or None. This fuzzy matching logic enables the system to maintain multiple competing hypotheses simultaneously, preventing premature commitment to a single, potentially incorrect interpretation of enemy intent.

Key Insights and Results

The PRM was implemented on a Symbolics Lisp Machine and demonstrated several key capabilities:

  • Adversarial Reasoning: The model explicitly accounts for opponents who may attempt to hide their intentions.
  • Multi-Hypothesis Tracking: The system successfully updated multiple intent hypotheses in real-time, providing users with a dynamic tactical scenario overview.
  • Heuristic Efficiency: The best-first search strategy effectively managed the computational load of high-speed tactical data.
  • Temporal Flexibility: By allowing for "relaxed" parameter boundaries, the model handles overlapping tactical events rather than treating them as discrete, isolated blocks.

Limitations and Future Directions

While pioneering, the 1986 PRM had notable limitations. It was restricted to analyzing a single threat agent at a time, with future work aimed at coordinating multi-agent recognition. The authors also identified the need for more robust uncertainty modeling (such as fuzzy logic) and the exploitation of parallel processing to meet the "seconds-to-live" requirements of actual combat. Finally, the initial model was static, pointing toward the necessity of machine learning to make the system truly adaptive to evolving enemy tactics.

Summary updated

BibTeX
@inproceedings{azarewicz1986plan,
  author = {Jerome Azarewicz and Glenn Fala and Ralph Fink and Christof Heithecker},
  title = {{Plan Recognition for Airborne Tactical Decision Making}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure