Introduction
In classical logic and game theory, researchers often focus on an agent's ability to "know" they have a strategy to achieve a specific goal. However, real-world autonomous systems—ranging from self-driving cars to missile defense platforms—do not operate on perfect knowledge. Instead, they operate on data provided by sensors. This paper introduces a formal logical framework for "doxastic strategies," which are strategies an agent believes will succeed based on data they trust, even if that data is fundamentally inaccurate.
The Problem: The Gap Between Knowledge and Belief
The authors highlight the critical distinction between knowing a strategy works (based on true information) and believing a strategy works (based on trusted, but potentially false, data). A tragic real-world example of this gap is the 1991 Patriot missile failure in Dhahran.
During this incident, the system "believed" it had a valid strategy to intercept an incoming Scud missile based on its radar data. However, a software clock drift caused the time data to be inaccurate. While the system trusted the data, the data was not trustworthy. Consequently, the system calculated a trajectory for a position where the Scud no longer was, leading to a fatal miss.
A New Logical Framework
To address this, the authors develop a modal logic that merges Epistemic Logic (the logic of knowledge) with Coalition Power (the logic of group achievement). The system is built around two primary new modalities:
- $B^T_X \phi$: This represents that, under the assumption that dataset $T$ is trustworthy, observing dataset $X$ informs the agent's belief that $\phi$ is true.
- $[C]^T_X \phi$: This represents that, under the assumption that dataset $T$ is trustworthy, observing dataset $X$ allows a coalition $C$ to identify a strategy to achieve $\phi$.
The logic is governed by nine core axioms, including Trust, Cooperation, and Strategic Introspection. A standout is the Trust Axiom ($B^T_X(B^T_Y \phi \to \phi)$), which captures the internal logic of an agent: while its beliefs might be wrong, the agent believes that if its data is trustworthy, its conclusions will be correct.
Formal Semantics and Completeness
The authors define "Games" as the underlying model for this logic. In these models, states are connected by indistinguishability relations based on specific data variables, and transitions are governed by the actions of agents. This approach links what an agent knows directly to the sensors (variables) it has access to, providing a more realistic model for AI than traditional abstract indistinguishability.
The mathematical core of the paper is a proof of "strong completeness." This proof demonstrates that the proposed axioms perfectly capture all valid statements within the defined game semantics, ensuring the system is both sound (it proves no falsehoods) and exhaustive.
Key Insights and Results
The primary contribution of this work is the formalization of "believe-how" as a replacement for "know-how" in autonomous systems. By modeling the Patriot failure with the formula $[Patriot]^T_X (\text{"Scud is destroyed"})$, the authors show that the formula can evaluate to true in the agent's belief state while remaining false in the actual state of the world. This precisely captures the logic of the error that led to the 1991 disaster.
Future Directions
While the current framework provides a robust foundation for one-shot strategic games, the authors identify several areas for expansion:
- Extensive Games: Moving from one-shot decisions to multi-step tracking and interaction.
- Non-Binary Trust: Moving beyond "trustworthy vs. untrustworthy" to incorporate probabilistic or fuzzy trust levels to better handle sensor noise.
- Dynamic Logic: Integrating belief revision to model how an agent's strategy changes as new data arrives over time.
Summary updated