Accounting for Context in Plan Recognition, with Application to Traffic Monitoring

David V. Pynadath, Michael P. WellmanUAI 1995 · OpenReview

TL;DR

This paper introduces a general Bayesian framework for plan recognition that explicitly models the agent's mental state and the context in which plans are generated. By incorporating this context, the authors demonstrate improved plan recognition accuracy in a traffic monitoring scenario, showing how understanding the driver's goals and beliefs leads to better predictions of their actions.

What Problem Does This Paper Address?

Plan recognition, the task of inferring an agent's underlying plan from its observed actions, is crucial for predicting future behavior, understanding past actions, and influencing an agent's decisions. Traditional approaches often focus solely on the relationship between a plan and its observable effects. However, this paper argues that a more comprehensive view is needed, one that accounts for the context in which the plan was generated. This context includes the agent's mental state (beliefs, preferences, capabilities) and the surrounding environmental factors that influence plan selection. The authors aim to bridge this gap by developing a framework that leverages this contextual information for more robust plan recognition, using traffic monitoring as a concrete application.

Methods

The core of the paper is a general Bayesian framework for plan recognition. This framework is structured around a causal model of planning, encompassing:

  • Context: Observable environmental factors influencing the agent.
  • Mental State: The agent's beliefs, preferences, and capabilities.
  • Planning Process: How the agent selects a plan based on its mental state and context.
  • Activity: The actions the agent performs to execute its plan.
  • Effects: Changes in the world state resulting from the agent's actions.

This causal model is translated into a Bayesian network. The authors emphasize that this framework is designed to be general, allowing for various levels of specificity in representing plans and planning processes. For their traffic monitoring application, they construct a detailed Bayesian network that models a driver's decision-making process, including factors like lane position, speed, presence of other vehicles, and desired exit. They highlight the novelty of explicitly modeling the plan generation context, rather than just the plan execution, and how this context influences the probability of different plans.

Datasets

Not discussed in the paper. The paper focuses on a simulated traffic monitoring domain and does not describe the use of real-world datasets.

Key Insights and Ideas

  • Context is King (for Plan Recognition): The most significant insight is that understanding the why behind a plan (the agent's mental state and environmental context) is as crucial as observing the what (the actions). This allows for more accurate disambiguation of ambiguous actions.
  • Bayesian Networks as a Unifying Framework: The paper demonstrates the power of Bayesian networks to integrate diverse aspects of plan recognition—context, mental state, planning, and execution—into a single, coherent probabilistic model.
  • Rationality as a Constraint: The assumption of agent rationality (i.e., agents choose plans that best satisfy their preferences given their beliefs) is a powerful constraint that can significantly prune the space of possible plans.
  • Traffic Monitoring as a Testbed: The traffic domain is a good example because driver actions (lane changes, acceleration) are often ambiguous without considering context like surrounding vehicles, desired speed, and destination.

Key Results

The paper demonstrates the effectiveness of their context-aware Bayesian framework through a traffic monitoring example. While specific quantitative results like accuracy percentages are not the primary focus, the paper illustrates how incorporating contextual observations (e.g., the presence of a car to the left) can significantly alter the posterior probabilities of different driver plans. For instance, observing a blocked left lane increases the probability of a "passing" maneuver, whereas without this context, a simple "lane change" might be predicted. The framework allows for probabilistic inference, providing not just a single predicted plan but a distribution of probabilities over possible plans and future actions.

Limitations and Future Work

  • Generality and Scalability: The authors acknowledge that the generality and scalability of their framework remain to be seen. While the traffic example is specific, they believe the general structure is applicable.
  • Complex Preferences and Decision Processes: The driver model had simplified goals (exit, target speed). Real-world agents may have more complex preferences and decision-making processes that are harder to capture.
  • Communication: The paper notes that agent communication is largely unexplored within their model. Future work aims to incorporate communication acts as actions themselves.
  • Real-world Data: The current work relies on a simulated domain. Applying the framework to real-world data would be a significant next step.
  • Sensor Noise: While mentioned as an extension, the current model assumes perfect sensors. Incorporating sensor noise would add realism.

Summary updated

Abstract

Typical approaches to plan recognition start from a representation of an agent's possible plans, and reason evidentially from observations of the agent's actions to assess the plausibility of the various candidates. A more expansive view of the task (consistent with some prior work) accounts for the context in which the plan was generated, the mental state and planning process of the agent, and consequences of the agent's actions in the world. We present a general Bayesian framework encompassing this view, and focus on how context can be exploited in plan recognition. We demonstrate the approach on a problem in traffic monitoring, where the objective is to induce the plan of the driver from observation of vehicle movements. Starting from a model of how the driver generates plans, we show how the highway context can appropriately influence the recognizer's interpretation of observed driver behavior.
BibTeX
@inproceedings{pynadath1995accounting,
  author = {David V. Pynadath and Michael P. Wellman},
  title = {{Accounting for Context in Plan Recognition, with Application to Traffic Monitoring}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1995},
  url = {https://openreview.net/forum?id=4H3YncnkzH},
}
Figure