Is Everything Going According to Plan? Expectations in Goal Reasoning Agents

Héctor Muñoz-Avila, Dustin Dannenhauer, Noah ReifsnyderAAAI 2019 · DOI

Abstract

In autonomous systems, "Goal Reasoning" allows agents to dynamically formulate and revise their objectives. However, effective goal revision requires the agent to first detect discrepancies between its internal model and the real world. This paper provides a comprehensive taxonomy of "expectations"—the internal benchmarks agents use to monitor plan success. By categorizing six distinct forms of expectations and evaluating them against domain features like non-determinism and hierarchical planning, we identify critical trade-offs in sensing costs and highlight significant gaps in current AI literature.

Introduction

In the field of AI safety and autonomy, agents must do more than follow static scripts; they must reason about their goals in real-time. A central challenge in this process is defining what constitutes a failure. Is it enough for an agent to check if its last action succeeded, or must it continuously verify that the final goal remains reachable? This paper addresses the lack of a unified framework for "expectations," which are the criteria an agent uses to detect discrepancies between its predicted state and the actual environment.

Figure 1: State transition diagram illustrating non-deterministic outcomes and the role of expectations in monitoring plan execution.

A Taxonomy of Expectations

The authors identify six distinct forms of expectations found in AI literature, each offering different levels of rigor and computational cost:

  1. Immediate Expectations: The simplest form, where the agent checks if the most recent action's effects are visible and if the next action's preconditions are met.
  2. State Expectations: The agent projects the entire world state forward for every step, comparing the entire observed environment to its prediction.
  3. Informed Expectations: The agent tracks only the accumulated changes (the "delta") caused by its actions, which is particularly useful when the full initial state is unknown.
  4. Goal Regression Expectations: Working backward from the goal, the agent identifies the "weakest preconditions"—the bare minimum requirements at each step to ensure the final goal remains reachable.
  5. Goldilocks Expectations: A hybrid approach designed for Hierarchical Task Network (HTN) planning. It combines forward-projected effects with backward-regressed conditions to find a "just right" set of expectations when explicit goals are not clearly defined.
  6. Metacognitive Expectations: A high-level monitor that evaluates the agent's own reasoning process, asking whether the internal explanation for a failure is logically sound.

Key Insights and Analysis

The Sensing Trade-off

There is a direct relationship between the complexity of an expectation and its "sensing cost." While State Expectations provide high confidence, they require the agent to sense every variable in the environment. Conversely, Goal Regression is more efficient, requiring the agent to sense only the specific variables necessary to reach the goal.

The HTN Gap and Success Guarantees

Hierarchical Task Networks (HTNs) are popular for mimicking human-like task decomposition but often lack explicit goals. The Goldilocks approach bridges this gap by deriving expectations directly from the tasks. Crucially, the authors note that while "Immediate Expectations" are easy to compute, they do not guarantee goal achievement. Only State, Regression, Informed, and Goldilocks expectations provide a formal guarantee that execution will terminate in a successful state.

Results and Discussion

The primary contribution of this work is the mapping of expectation types against various domain features, such as non-determinism (ND), numeric conditions, and ontological reasoning.

Table 1: Comparison of expectation types across different domain features and planning formalisms, highlighting research gaps.

As shown in Table 1, while State Expectations are the most common, they are also the most expensive. Goal Regression is highly efficient but faces implementation hurdles in non-deterministic or hierarchical domains. The "question marks" in the table identify significant frontiers for future research, particularly regarding how to handle Numeric and Ontological data within Regression and Goldilocks frameworks.

Conclusion and Future Work

While this taxonomy provides a map for current research, several challenges remain. Most current models assume discrete, sequential actions and struggle with continuous or concurrent processes. Furthermore, most expectations currently require manual coding by experts. A major future direction for the field is enabling agents to autonomously learn what they should expect through environmental interaction and experience.

Summary updated

BibTeX
@inproceedings{munozavila2019is,
  author = {Héctor Muñoz-Avila and Dustin Dannenhauer and Noah Reifsnyder},
  title = {{Is Everything Going According to Plan? Expectations in Goal Reasoning Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019823},
}
Figure