Introduction
In their seminal 1987 work, Philip Cohen and Hector Levesque established a formal logical foundation for the concept of "intention" within artificial intelligence and multi-agent systems. Moving beyond the then-prevalent view of intentions as mere plans, the authors proposed that intention is a complex mental state characterized by a persistent commitment to a chosen goal. This framework provides a rigorous basis for modeling how rational agents maintain consistency, reason about side effects, and respond to plan failures.
The Problem: Plans vs. Intentions
Prior to this work, AI research often conflated an agent's intentions with its current plans—sequences of actions designed to achieve a state. Cohen and Levesque argued that this approach failed to capture two critical aspects of rational behavior:
- Persistence: A simple plan-executor might abandon a goal the moment an action fails. For example, if an agent plans to buy milk and finds the store closed, a plan-based model might simply terminate. A rational agent with an intention, however, should persist, perhaps by seeking an alternative store.
- The Side-Effect Problem: If an agent intends to cure a headache by taking aspirin, knowing it will cause an upset stomach, does the agent intend to have an upset stomach? Traditional models struggled to distinguish between intended outcomes and foreseen but undesired consequences.
Formalizing Rational Agency
The authors utilized modal logic and possible-worlds semantics to build their framework, incorporating several key operators:
- Belief (BEL) and Goal (GOAL): Fundamental modal operators representing an agent's internal state.
- Dynamic Logic: Used to model actions and their sequences over time.
- Persistent Goal (P-GOAL): The paper’s central innovation. A P-GOAL is defined as a goal that an agent will not relinquish until they believe it has been satisfied or have concluded it is impossible to achieve.
Under this logic, "intending an action" is defined as having a persistent goal to have performed that action, coupled with the belief that one is about to perform it.
Key Insights: Choice and Commitment
The framework posits that intention is the result of a choice (selecting a goal from various desires) and a commitment (the decision to persist with that goal over time).
The Screen of Admissibility
Intentions serve as a "screen of admissibility" for future reasoning. Once an agent adopts an intention, it constrains their future deliberations; they will not adopt new intentions that they believe would conflict with or render their existing commitments impossible.
Solving the Side-Effect Problem
The P-GOAL definition elegantly handles side effects. Because intention is tied to commitment, an agent only "intends" a consequence if they would "track" its success. If a headache is cured without the stomach ache, the agent is satisfied. If they truly intended the stomach ache, they would feel they had failed and seek to upset their stomach through other means.
Avoiding the "Little Nell" Paradox
The logic addresses the "Little Nell" paradox, where an agent might drop a goal because they believe it will succeed (e.g., "I am currently saving Nell, therefore she is safe, therefore I no longer need to save her"). By formalizing the temporal aspects of goals and beliefs, Cohen and Levesque ensure that agents maintain their commitments until the goal state is actually realized.
Theoretical Results
The paper provides formal theorems demonstrating the rationality of the proposed logic:
- Theorem 1 (Persistence to Eventuality): A competent agent with a persistent goal will eventually achieve it, provided it remains possible.
- Theorem 2 (Consistency): Agents cannot intentionally act to undermine their own persistent goals.
- Theorem 3 (Tracking): If an action fails but the goal remains achievable, the agent maintains the intention and will attempt to rectify the failure.
Limitations and Legacy
While foundational, the model acknowledges certain limitations. It assumes "logical omniscience," where agents automatically believe all logical consequences of their beliefs—a condition computationally infeasible for real-world systems. Furthermore, the paper focuses on internal commitment; "social commitment" (promises between agents) is identified as a necessary future layer.
Despite these limitations, Cohen and Levesque’s work remains a cornerstone of AI, providing the vocabulary and logical rigor necessary to discuss the mental states of autonomous agents.
Summary updated