Plan Recognition

Plan recognition is fundamentally an inverse planning problem. While automated planning generates a sequence of actions to achieve a goal, plan recognition works backwards from observed actions to hypothesize the most likely goal and plan. This is typically framed as a probabilistic inference task, often using Bayesian reasoning to evaluate the posterior probability of different plans given the observations and a model of the agent's planning process.

Classic examples illustrate the ambiguity inherent in observation.

• Keypad Problem: Observing the sequence [1, 4, 7] on a telephone keypad is consistent with multiple goals: dialing the number 147, typing the word 'GHP' using multi-tap, or even a failed attempt to press 8 (located at the center of 1, 4, and 7).
• Library Domain: An agent is observed moving to a room, picking up a book, and going to a checkout desk. The goal could be to borrow the book, return the book, or check a reference. The true plan is under-specified by the actions alone, requiring context to disambiguate.

Two primary computational approaches exist:

• Plan-Library Approach: Compares observations against a pre-enumerated library of known plans. It is efficient but limited to recognized scenarios. Used in script recognition and structured environments.
• Plan-Generation Approach: Uses a generative model of the agent's planning (e.g., a planner simulator) to dynamically generate candidate plans that explain the observations. This is more flexible and can handle novel situations but is computationally more expensive.

Algorithms rely on core assumptions about the observed agent:

• Rationality Assumption: The agent is executing a plan that is approximately optimal or rational with respect to its goals. This allows the recognizer to prune implausible, highly inefficient plans.
• Plan Cost: Many algorithms incorporate a cost metric for actions (e.g., time, effort). They infer that the agent is likely pursuing a goal that justifies the incurred cost, and that cheaper plans explaining the same observations are more probable.

Plan recognition is closely related but distinct from intent recognition. Intent recognition aims to identify the high-level goal or desire (the why), such as 'want to drink coffee'. Plan recognition infers the specific sequence of actions intended to achieve that goal (the how), such as 'walk to kitchen, grind beans, brew pot'. In integrated systems, plan recognition often serves as the mechanism to operationalize a recognized intent.

Plan recognition enables proactive systems:

• Assistive Technology: Smart homes can predict a resident's goal (e.g., preparing a meal) and automate subsequent steps (preheat oven) or offer help if a step is missed.
• Security & Monitoring: Identifying anomalous sequences of actions in network logs (port scan, vulnerability probe, data exfiltration) that constitute an attack plan.
• Human-Computer Interaction: Anticipating user needs in software interfaces by recognizing task patterns, enabling shortcut suggestions or automated macro completion.

The computational process of inferring the immediate goal or purpose behind an agent's observed action or communication. While plan recognition focuses on multi-step sequences, intent recognition often targets the objective of a single utterance or atomic action.

• Key Differentiator: Typically operates on a shorter time horizon than full plan recognition.
• Example: Classifying a user's query "turn on the lights" as an intent to ACTUATE_DEVICE rather than inferring a full evening routine plan.

A Bayesian reasoning framework for inferring an agent's hidden goals and beliefs by treating them as the cause of observed actions. It works backwards from actions to goals, assuming the observed agent is approximately rational.

• Core Assumption: The observed agent is executing a plan generated by a forward planning algorithm.
• Mechanism: Computes the posterior probability of goals P(Goal | Actions, World Model).
• Application: Foundational for model-based plan recognition in robotics and cognitive science.

A dominant software architecture for intelligent agents that structures decision-making around three key mental state components: Beliefs (world model), Desires (potential goals), and Intentions (committed plans). Plan recognition in BDI systems involves inferring this triad from behavior.

• Architectural Impact: Provides a formal framework for representing the mental states that plan recognition aims to uncover.
• Relation: Recognized plans map directly to inferred Intentions within the BDI model.

A computational approach where an agent models not only the world but also the mental models of other agents, which may include those agents' models of others. This creates nested reasoning essential for strategic interaction.

• Direct Link: Effective plan recognition in adversarial or cooperative multi-agent settings often requires recursive modeling (e.g., "I think my opponent is trying to make me think she is executing Plan A, so she must actually be executing Plan B").
• Orders of Theory of Mind: First-order (modeling others' beliefs), second-order (modeling others' beliefs about beliefs), etc.

The process of making decisions by explicitly modeling the likely decisions of other rational agents who are simultaneously modeling you. It extends plan recognition into interactive, game-theoretic contexts.

• Key Concept: Common Knowledge and Mutual Belief are critical for aligning strategic models.
• Application: Used in automated negotiation, poker-playing AI, and economic mechanism design. It requires plan recognition as a sub-process to predict others' actions.

The application of Theory of Mind capabilities—including plan recognition—in competitive or zero-sum scenarios to anticipate, deceive, and counter an opponent's strategies. It focuses on modeling hostile or self-interested intent.

• Contrast with Cooperative Recognition: Must account for active deception and information hiding.
• Techniques: Often involves generating and recognizing plans that include misleading actions or feints.
• Domain: Cybersecurity (modeling attacker campaigns), military simulation, and competitive games.