Revising Beliefs and Intentions in Stochastic Environments

Nima Motamed, Natasha Alechina, Mehdi Dastani, Dragan DoderIJCAI 2024

Abstract

In classical artificial intelligence, the Belief-Desire-Intention (BDI) architecture provides a robust framework for modeling agent behavior. However, traditional models often struggle with stochasticity—environments where actions have uncertain outcomes. This paper introduces a formal framework for the joint revision of beliefs and intentions in such environments. By leveraging a decidable probabilistic temporal logic (PLBP+), the authors provide representation theorems that characterize how rational agents should update their goals and worldviews when faced with probabilistic uncertainty.

Introduction: The Coherence Problem

Most existing models for intention revision assume a deterministic world where an action always leads to a predictable result. Real-world environments are rarely so accommodating; for instance, an agent attempting to purchase a concert ticket may succeed or find the event sold out. This inherent uncertainty creates a "coherence" problem. It is insufficient for an agent to maintain consistent beliefs and consistent intentions in isolation; the two must cohere. It is fundamentally irrational to intend to achieve a goal if one believes the probability of success is negligible.

This research tackles the challenge of defining how agents should update these intertwined states in a world governed by probabilities, ensuring that intentions remain grounded in the agent's evolving understanding of the environment.

The PLBP+ Framework

The authors build their framework on PLBP+ (Probabilistic Logic of Bounded Policies). This logic allows agents to reason about $n$-step strategies and the probability of various outcomes over time. A key methodological contribution is the introduction of $\theta$-Coherence, a novel definition of joint consistency. Under this definition, an intention is only considered coherent with a belief if there exists a policy that achieves the intention with a probability higher than a chosen threshold $\theta$.

By parameterizing the logic with $\theta$, the framework accommodates different agent "personalities." A risk-averse agent might require a high $\theta$ (near certainty) to form an intention, while a more adventurous agent might be willing to gamble on intentions with lower success probabilities.

Strong vs. Weak Beliefs

A critical distinction made in the paper is between "strong" and "weak" beliefs:

  • Strong Beliefs: These represent what the agent knows about the environment's mechanics and current state.
  • Weak Beliefs: These represent what the agent expects to happen given its current intentions.

A key insight of the paper is that intentions shape beliefs. If an agent revises an intention—for example, deciding to save money rather than attend a concert—it must also revise its weak beliefs. It should no longer expect to be at the concert venue the following evening. The paper provides a unified operator that first updates world beliefs and then re-evaluates intentions to maintain this internal harmony.

Rationality Postulates and Representation Theorems

To ground the framework in formal logic, the authors adapt the classic AGM (Alchourrón, Gärdenfors, and Makinson) and Katsuno & Mendelzon postulates for belief revision. These postulates are re-engineered to function within probabilistic models, specifically Pointed Markov Decision Processes (pMDPs).

The authors prove several landmark representation theorems (Theorems 2, 3, and 4). These theorems demonstrate that any revision operator satisfying their defined "rationality" axioms corresponds to a semantic ordering of "closest" possible worlds. This provides a mathematical guarantee that if an agent follows these rational rules, its behavior is equivalent to selecting the most similar possible world that accommodates the new information.

Computational Feasibility

Probabilistic logics often face challenges with infinite state spaces due to the continuous nature of probability. The authors overcome this by introducing Finitely Expressible and Faithful (FEF) assignments. This approach partitions infinite spaces into a finite number of "levels," making the revision process computable. The paper proves that checking the satisfiability of PLBP+ formulas is decidable (Theorem 1), with revision operators computable in 2EXPSPACE.

Conclusion and Future Directions

This work establishes the first formal bridge between probabilistic logic and the joint revision of beliefs and intentions. While the current framework focuses on single-step revisions, it lays the groundwork for several future avenues of research:

  • Iterated Revision: Exploring how agents change their minds over a long sequence of conflicting observations.
  • Success Constraints: Investigating "consistent AGM revision," where an agent might reject new intentions if they are fundamentally incompatible with core beliefs.
  • Plausibility Nuance: Developing more granular ways of ranking likely worlds beyond the current level-based approach.

By providing a rigorous logical foundation for agent behavior in uncertain worlds, this research moves the BDI architecture closer to the complexities of real-world application.

Summary updated

BibTeX
@inproceedings{motamed2024revising,
  author = {Nima Motamed and Natasha Alechina and Mehdi Dastani and Dragan Doder},
  title = {{Revising Beliefs and Intentions in Stochastic Environments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure