The Model-Based Approach to Autonomous Behavior: A Personal View

Hector GeffnerAAAI 2010 · DOI

Introduction

The fundamental challenge in autonomous Artificial Intelligence can be distilled into a single, deceptively simple question: What action should the agent take next? In his synthesis of the field, Hector Geffner explores the evolution of "model-based" AI, positioning planning not merely as a search for paths, but as the central engine for intelligent behavior. Drawing on Karl Popper’s concept of "Popperian creatures"—agents that use models to let their hypotheses die in their stead—Geffner argues that the future of autonomy lies in our ability to efficiently solve, translate, and learn from world models.

A Taxonomy of Autonomous Behavior

Geffner categorizes the history of AI into three distinct approaches to the problem of action selection:

  1. Programming-based: Human experts write the logic and control rules. While computationally efficient, this approach is brittle and fails to generalize to unforeseen environments.
  2. Learning-based: Agents derive behavior from experience, typically through Reinforcement Learning (RL). While flexible, these methods often struggle with unobservable states, sparse rewards, and the sheer computational cost of scaling to complex domains.
  3. Model-based (Planning): The agent is provided with a formal model of the world—including actions, sensors, and goals—and must compute a solution. This is the most general approach, but it faces the significant hurdle of computational intractability.

The Evolution of Classical Planning

For decades, classical planning was viewed as a niche area of logic. However, the field has undergone a revolution, moving from solving "toy" puzzles to handling complex problems with hundreds of variables. This breakthrough was not driven by hardware alone, but by the discovery of automatic heuristic extraction.

Modern planners like FF and LAMA do not require humans to provide domain-specific hints. Instead, they analyze the propositional structure of the problem description to derive heuristics—such as "landmarks" (fluents that must be achieved) or "helpful actions"—that prune the search space. This allows for "satisficing" solutions to be found almost instantly in deterministic, fully observable environments.

Beyond Determinism: Translation and Compilation

Real-world autonomy rarely offers the luxury of full observability or deterministic outcomes. Traditionally, these problems were handled by specialized solvers for Markov Decision Processes (MDPs) or Partially Observable MDPs (POMDPs). Geffner highlights a provocative alternative: Translation-based Planning.

Instead of building "native" solvers for every type of uncertainty, researchers are increasingly "compiling away" complexity into standard classical planning formats. For example, soft goals, LTL constraints, and even certain types of uncertainty (using a "width" parameter) can be translated into a classical model. This allows the field to leverage highly optimized classical solvers for non-classical problems.

This approach has led to the "Replanning Paradox": in many stochastic environments, simple classical planners that ignore most outcomes and simply "replan" when the environment deviates from the model often outperform complex, native MDP solvers that attempt to account for every possible contingency upfront.

Key Insights and Future Directions

Geffner identifies several frontiers that will define the next generation of autonomous systems:

  • Learning from Search: Just as SAT (Satisfiability) solvers have improved by "learning" from backtracks to avoid future mistakes, planners should evolve to adjust their heuristic functions based on search failures.
  • Automatic Abstraction: Humans naturally organize tasks into hierarchies (e.g., "go to the airport" involves many sub-tasks). A major open challenge is the automatic compilation of detailed world models into these abstract Hierarchical Task Networks (HTNs).
  • The Wumpus Challenge: Despite progress, domain-independent solvers still struggle with the "Wumpus World"—a benchmark requiring a combination of partial observability and complex logic at scale.
  • Closing the Loop: The ultimate goal is to bridge the gap between learning and planning. This involves learning the world model itself from experience and then using that model for planning, effectively combining the flexibility of RL with the reasoning power of classical AI.

Conclusion

The model-based approach provides a robust framework for understanding and engineering autonomous behavior. By focusing on the structural properties of problems and refining the tools for translation and search, the field is moving toward agents that can reason through uncertainty and adapt to complex, multi-agent environments without the need for exhaustive manual programming.

Summary updated

BibTeX
@inproceedings{geffner2010model,
  author = {Hector Geffner},
  title = {{The Model-Based Approach to Autonomous Behavior: A Personal View}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7765},
}
Figure