Author: Charles L. Ortiz, Jr. (1993)
Abstract
Formalizing the concept of "prevention" is a fundamental challenge in Artificial Intelligence, particularly for autonomous agents operating in complex environments. Traditional logical definitions often equate prevention with making an event physically impossible—a standard that is frequently unattainable and linguistically inaccurate. This paper proposes an "agent-centered" semantics for prevention. By integrating Belief-Desire-Intention (BDI) logic with branching temporal operators and contextual lattices, the author provides a framework where prevention is defined by an agent's mental states and situational constraints rather than absolute physical impossibility.
1. The Problem with "Necessary Prevention"
In classical AI logic, prevention is typically defined as "necessary prevention": an event $e$ prevents $e'$ if $e'$ was possible before $e$ occurred but became impossible immediately after. Ortiz identifies two primary flaws in this "naive" view:
- The Planning Problem: In a complex world, an agent can rarely guarantee that an event is absolutely impossible across all future timelines. If absolute impossibility were the requirement, agents could almost never claim to have prevented anything.
- The Linguistic Problem: Human language often uses "prevented" in contexts where the action remains physically possible but is constrained by other factors. For example, a red light prevents a driver from turning, even though the driver is physically capable of performing the maneuver.
2. An Agent-Centered Formalism
To address these issues, the paper builds upon the BDI (Belief, Desire, Intention) framework. Ortiz introduces several logical extensions to better capture the nuances of human-like reasoning:
- Branching Modal Operators: These allow for reasoning across multiple potential futures rather than a single linear path.
- Event Typing: A
type(e1, e2)predicate is introduced to recognize that a single physical act (e.g., flipping a switch) can have multiple semantic descriptions (e.g., illuminating a room). - Agent-Centered Definition (PR1): This core definition requires that the "patient" (the agent being prevented) must have a specific mental state—an intention to perform an action—which the "preventer" then disrupts.
3. Intentionality and the "Screen of Admissibility"
A key insight of the paper is that prevention is inherently tied to an agent's goals. Ortiz introduces the concept of a "screen of admissibility," where an agent's intentions act as a filter for their future actions. If an agent intends to reach a goal within a certain timeframe, they will ignore paths that are too slow or inefficient.
As illustrated in Figure 1, an agent (the "Amazon" in the paper's thought experiment) navigating from point A to point S can be "prevented" from taking a specific route not just by physical barriers, but by changing the agent's beliefs or the environmental context. For instance, placing a minor obstacle that makes a path inconsistent with the agent's time-sensitive goals effectively prevents that path from being taken, even if it remains physically traversable.
4. Contextual Impossibility and Lattices
Ortiz proposes the use of contextual lattices ($C_i$) to handle the relativity of possibility. These are sets of possible worlds restricted by specific rules, such as the laws of physics, legal codes, or social norms. This allows an AI to reason that an action is "impossible" within a specific context (e.g., "illegal") even if it is achievable in the broader physical world. This contextual approach aligns AI reasoning more closely with human social and legal structures.
5. Conclusion and Future Work
The primary contribution of this work is a formal, symbolic definition of agent-centered prevention that accounts for side-effect prevention and contextual constraints. While these strategies are "good bets" rather than "sure bets"—relying on the assumption that the patient agent is rational and possesses certain beliefs—they provide a much more practical foundation for AI planning. Future work remains in addressing the complexity of counterfactual reasoning and defining how agents dynamically switch between different contextual micro-worlds during active reasoning.
Summary updated