Introduction
Communication in high-stakes environments like Air Traffic Control (ATC) is rarely a simple dyadic exchange. Instead, it occurs over shared radio frequencies where multiple pilots and controllers listen in, building situational awareness from conversations they are not directly part of. Novick and Ward address the challenge of modeling these multi-party, task-oriented dialogues, focusing on how agents track "mutuality of belief"—the distinction between what an agent knows privately and what they believe is shared knowledge among others.
Multi-party Belief Tracking
The core of the model is a triplet representation: belief(proposition, truth-value, belief-groups). The "belief-groups" parameter identifies which sets of agents mutually hold a specific belief. This allows the system to represent complex states, such as a controller believing that a pilot has confirmed an altitude instruction, thereby establishing it as mutual knowledge.
As illustrated in Figure 1, the model explicitly accounts for the "overhearer" (Agent C). While Agent C does not participate in the collaborative grounding process between A and B, they still update their mental model based on the observed "utterance." By applying the same mutuality rules as the active participants, the overhearer successfully maintains situational awareness of the shared environment.
The saso Simulation and Rule Architecture
The model was implemented using saso, a Prolog-based rule-based shell designed for multi-agent problems. The simulation logic is categorized into four distinct rule sets that govern how agents perceive, predict, plan, and act within the Instrument Landing System (ILS) approach procedure.
As detailed in Figure 2, the architecture relies on:
- Mutuality Rules: Converting perceived acts into internal beliefs and updating the mutuality of information.
- Expectation Rules: Allowing agents to anticipate responses, such as expecting an acknowledgment following a directive.
- Planning Rules: Decomposing high-level goals (e.g., authorizing an approach) into actionable subgoals.
- Performative Rules: Executing intentions as specific speech acts like
acknowledge,authorize, orreport.
Evaluation and Results
The model was validated against a specialized corpus of four complete ATC dialogues recorded at Portland International Airport. Despite the small sample size, these high-fidelity protocols allowed for a granular, act-for-act comparison between the simulation and real-world behavior.
The simulations were highly successful, matching the original transcripts on an act-for-act basis and correctly tracking the internal belief states of both active participants and overhearers. Notably, the model remained robust even when dialogues were interleaved, mimicking the "busy frequency" environment of a real control tower where multiple conversations occur simultaneously.
Discussion and Limitations
While the model successfully demonstrates the mechanics of mutual belief in multi-party settings, several limitations remain. The saso environment assumes perfect radio transmission and immediate execution of instructions, ignoring the complexities of signal noise or delayed tasks. Furthermore, the current planning rules are domain-specific to the ILS task. Future work aims to integrate world-event simulations—such as weather changes or aircraft performance—and more generalized planning mechanisms to handle broader communicative contexts beyond the ATC domain.
Summary updated