Abstract
Ernest Davis's 1988 seminal work establishes a formal logical framework for reasoning about an agent's ignorance based on the physical constraints of visual perception. By axiomatizing the "locality" of vision—the principle that physical barriers like walls limit what an agent can know—the paper enables AI systems to logically deduce the absence of knowledge from spatial configurations.
Introduction: The Problem of Proving Ignorance
In early AI research, the focus was primarily on knowledge acquisition. However, reasoning about what an agent does not know is equally critical, especially for planning secrets or predicting another agent's behavior. Proving "not knowing" is notoriously difficult in standard logic without a formal model of the physical environment. For instance, how can an AI formally prove that Claire, standing behind a windowless wall, is ignorant of events occurring on the other side? Davis addresses this by bridging high-level epistemology with the low-level physics of vision, such as occlusion and line-of-sight.
Methods: A Formal Logic of Vision
Davis introduces a "Toy World" governed by simplified physics—rigid objects and continuous motion—and develops a formal language to describe it.
Layouts vs. Situations
The framework distinguishes between a layout (a static physical snapshot) and a situation (a layout paired with an agent's knowledge state).
V-Compatibility and Axioms
Central to the theory is the "visual compatibility" (V-compatibility) relation. Two layouts are V-compatible if they appear identical to an agent due to occlusion. Davis proposes eight core axioms (A.1–A.8) to govern this space. While most cover standard epistemic logic (memory, internal clocks), the "Ignorance Axiom" (A.8) is the most significant. It posits that if a course of events is physically possible and consistent with an agent's past and present perceptions, the agent cannot know those events did not occur.
Key Insights: Occlusion and Envelopes
The paper's most profound insight is that ignorance is a logical consequence of optical and geometric laws. Davis formalizes the concept of an envelope—a region of space hidden from an agent's view. By defining these envelopes, one can formally prove that an agent is ignorant of any activity within that region.
Key Results
The formal model supports five types of qualitative inferences:
- Inferring Ignorance: Proving Claire is unaware of Steve's actions behind a wall.
- Gaining Knowledge: Deducing the absence of an object (e.g., no cows in an office) by observing an empty space.
- Planning for Perception: Knowing that moving to a new location will reveal hidden information.
- Temporal Inference: Using continuous observation to track an object's location over time.
- Nested Knowledge: Reasoning about what another agent knows based on their line of sight.
Limitations and Future Work
The framework faces several challenges. Axiom A.8 is arguably too strong, leading to the "Invisible Bee" problem: an agent can never be certain of anything outside their immediate view because "physically possible" but unobserved events cannot be logically ruled out. Additionally, the model assumes "logical omniscience," where agents instantly know all consequences of their perceptions—a feat impossible for real-world robotics. Davis suggests that future iterations should move toward "plausible reasoning" to mitigate these issues.
Summary updated