Self Monitoring, Goal Driven Autonomy Agents

Dustin DannenhauerIJCAI 2016

Abstract

In dynamic and unpredictable environments, intelligent agents must go beyond simple path correction to maintain effectiveness. This paper explores the Goal-Driven Autonomy (GDA) framework, which enables agents to monitor their own performance and the environment to detect "discrepancies" between expectations and reality. Rather than merely recalculating a path to a static goal, GDA agents can autonomously decide to change their goals entirely when they detect significant anomalies. We examine technical implementations using Description Logic, Hierarchical Task Networks, and metacognitive monitoring across domains such as Real-Time Strategy (RTS) games and naval simulations.

Introduction

Traditional AI systems often rely on "replanning" when faced with environmental changes—finding a new sequence of actions to achieve a pre-defined objective. However, in complex scenarios like naval operations or RTS games, the original goal may become obsolete. For instance, a naval escort encountering an unexpected threat might need to pivot from its delivery mission to a reconnaissance or defensive mission. The challenge lies in discrepancy detection: identifying when reality diverges from expectations and determining whether a strategic shift in goals is required.

Methodology

The core of this research is the Goal-Driven Autonomy (GDA) cycle, which provides a structured approach for agents to manage their objectives dynamically.

The Goal-Driven Autonomy (GDA) conceptual model, illustrating the cycle of discrepancy detection, explanation, and goal formulation.

As shown in Figure 1, the GDA cycle consists of four primary stages:

  1. Discrepancy Detector: Compares real-world observations against the agent's expectations.
  2. Explanation Generator: Hypothesizes the underlying cause of a detected discrepancy (e.g., an enemy unit blocking a path).
  3. Goal Formulator: Proposes new goals based on these explanations.
  4. Goal Manager: Selects the most appropriate goal for the agent to pursue.

To enhance this cycle, three specific technical approaches are utilized:

  • Description Logic (DL) Rules: These define high-level expectations. In RTS games, DL allows agents to reason about abstract concepts like "territory control" rather than just low-level unit coordinates.
  • Informed Expectations via HTN: By leveraging Hierarchical Task Networks (HTN), agents generate expectations based on the predicted effects of their planned actions. This filters out environmental "noise," focusing only on changes that impact the current plan.
  • Metacognitive Monitoring: The agent maintains a "cognitive trace" of its internal processes. If an internal failure is detected (e.g., a planner stalling), the agent can autonomously swap components, such as switching to a different planning algorithm.

Evaluation and Results

The GDA framework was evaluated across several simulated environments:

  • StarCraft (RTS): Agents using inferred expectations demonstrated complex, coordinated tactics, such as multi-group attacks, which are typically difficult for standard bots.
  • Navy Strategy Simulation: This domain tested the agent's ability to balance convoy protection with threat identification in high-stakes scenarios.
  • Marsworld-like Domain: In this limited-perception environment, agents successfully chose between different signaling methods (beacons, flares, or smokefires) based on exploratory discoveries.

Key results indicate that goal flexibility is often more robust than path flexibility. By moving from unit-level to region-level expectations in RTS games, agents performed human-like tactics. Furthermore, metacognitive monitoring allowed agents to successfully self-correct by swapping planning faculties when internal failures were detected.

Discussion and Conclusion

The GDA framework shifts the focus from "how" to "what," allowing agents to adapt their very objectives to changing circumstances. By treating the agent's own cognitive processes as an object of study (metacognition), the system becomes more resilient to both external and internal failures.

Future work will address challenges in limited-perception environments, where generating grounded plans is difficult. Expanding metacognitive expectations to include memory and perception, as well as improving long-term autonomous learning, remains a priority for developing truly autonomous agents capable of operating in the most unpredictable environments.

Summary updated

BibTeX
@inproceedings{dannenhauer2016self,
  author = {Dustin Dannenhauer},
  title = {{Self Monitoring, Goal Driven Autonomy Agents}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure