What's the Context? Implicit and Explicit Assumptions in Model-Based Goal Recognition

Peta Masters, Mor VeredIJCAI 2021 · DOI

TL;DR

This paper argues that the success of model-based goal recognition (GR) hinges on a multitude of implicit and explicit assumptions about the environment, observations, agents, and goals. By systematically cataloging these assumptions and their common defaults, the authors provide a valuable framework for researchers to identify research gaps and for practitioners to select appropriate GR models.

What Problem Does This Paper Address?

Goal recognition (GR) is the AI problem of inferring an agent's intent from its observed behavior. While many approaches exist, model-based GR, which uses a model of the domain to infer goals, has gained traction. However, the authors observe that the field often overlooks the crucial role of assumptions underlying these models. These assumptions, ranging from the nature of the environment (e.g., dynamic vs. static) to the agent's behavior (e.g., rational vs. deceptive), significantly impact a GR system's performance and applicability. The paper aims to clarify these assumptions, highlight common defaults, and expose areas where relaxing these defaults can lead to novel research directions.

Methods

The paper's primary contribution is a comprehensive survey and categorization of assumptions in model-based GR. The authors meticulously review existing literature, identifying 20 distinct assumptions. They categorize these assumptions into four main groups: environment, observations, agents (actor and observer), and goals. For each assumption, they discuss its meaning, common default values (marked with an asterisk), and the implications of relaxing it. A key output is a table that maps various GR approaches to the assumptions they accommodate or relax, providing a visual overview of the research landscape. The authors emphasize the importance of explicitly stating these assumptions for clarity and reproducibility.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Assumptions are the bedrock of GR: The paper strongly emphasizes that every GR model is built upon a foundation of assumptions, and understanding these is critical for both development and application.
  • Default assumptions simplify but limit: The authors highlight that default assumptions often represent the simplest case (e.g., static, deterministic environments, rational agents), which makes implementation easier but less representative of real-world complexity.
  • The observer's perspective matters: GR assumptions are often framed from the observer's viewpoint, leading to concepts like "keyhole" (unaware observer) vs. "intended" (cooperative observer) recognition.
  • Confusion arises from implicit assumptions: Many assumptions are not explicitly stated, leading to potential misunderstandings and misinterpretations of research contributions. For example, the term "noise" is used ambiguously in the literature.
  • Relaxing assumptions drives progress: The paper demonstrates how researchers iteratively relax default assumptions to tackle more complex scenarios, leading to advancements in areas like dynamic goals, suboptimal behavior, and partial observability.
  • "Extended GR" flips the perspective: The concept of "Extended GR" is introduced, where GR is used not to find the agent's true goal, but to understand how a hypothetical observer might interpret the agent's actions, particularly relevant for human-robot interaction.

Key Results

The paper's main "result" is the systematic cataloging and analysis of assumptions in model-based GR, presented in detailed tables. While not presenting new empirical results, it provides a crucial meta-analysis of the field. The authors identify common default assumptions and highlight areas where research has begun to relax these defaults, such as:

  • Environment: Moving from static to dynamic and from deterministic to stochastic environments.
  • Observations: Incorporating partial observability, missing observations, and noisy data.
  • Agents: Addressing suboptimal behavior and the distinction between keyhole, intended, and deceptive agents.
  • Goals: Exploring multi-goals and dynamic goals.

The paper also points out specific contributions that have been pioneers in relaxing certain assumptions, such as Baker et al. [2009; 2011] for dynamic goals and Sohrabi et al. [2016] for incorporating noise and fluents.

Limitations and Future Work

The authors acknowledge that their list of assumptions is not exhaustive and was selected based on explicit mentions in the literature. They suggest future work could involve:

  • Expanding the assumption set: Incorporating assumptions from related fields like machine learning (e.g., independent vs. non-independent data) or network security.
  • Considering key modeling elements: Further analyzing challenges related to goals, observations, actors/observers, and environments, and identifying new assumptions based on real-world complexities (e.g., actors changing their minds, goals not being from a known set).
  • Combining research streams: The authors express interest in combining work on active goal recognition (where the observer can intervene) with counter-planning for dynamic environments.
  • Explicitly stating assumptions: A recurring theme is the call for researchers to be more explicit about their assumptions to improve clarity and facilitate future research.

Summary updated

BibTeX
@inproceedings{masters2021what,
  author = {Peta Masters and Mor Vered},
  title = {{What's the Context? Implicit and Explicit Assumptions in Model-Based Goal Recognition}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/615},
}
Figure