From Conditional Oughts to Qualitative Decision Theory

Judea PearlUAI 1993

TL;DR

This paper introduces a new framework for qualitative decision theory that extends classical deontic logic by incorporating causal relationships and order-of-magnitude reasoning for beliefs and utilities. This allows for a more robust analysis of conditional "ought" statements, enabling qualitative planning under uncertainty.

What Problem Does This Paper Address?

Traditional deontic logic struggles with conditional "ought" statements like "You ought to do A if C." These statements implicitly involve complex reasoning about actions, observations, beliefs, utilities, and causal consequences, which classical deontic logic, focused solely on logical implication, cannot adequately capture. Existing approaches often fall into paradoxes or lack the necessary formal machinery to handle the interplay of these modalities. The paper aims to bridge this gap by providing a decision-theoretic account that can reason about "ought" statements in a qualitative, yet rigorous, manner, especially in situations involving uncertainty.

Methods

The paper proposes an epistemic state defined by three components:

  1. Ordinal Belief Ranking (κ(ω)): This represents beliefs using non-negative integers, where lower values indicate higher surprise or lower probability. It's an order-of-magnitude abstraction of probability.
  2. Causal Network (Γ): A directed acyclic graph (DAG) that encodes causal relationships between variables, defining how actions influence beliefs and how observations update them.
  3. Integer-Valued Utility Ranking (μ(ω)): This assigns integer ranks to worlds based on their desirability, representing qualitative utilities (e.g., very desirable, bearable, very undesirable).

The core of the method lies in defining a criterion for the assertability of conditional "ought" statements, O(A|C), which states that "We ought to do A, given C" if the expected utility of doing A (after observing C) is substantially higher than the expected utility of not doing A. This involves:

  • Action Updates: Defining how performing an action A transforms the belief ranking κ(ω) into a post-action ranking κ_A(ω). This is handled by a persistence model that balances causal influences with the tendency of properties to persist.
  • Conditional Oughts: Formulating the assertability criterion using the post-action belief ranking and the utility ranking.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Causality is Crucial for "Oughts": The paper argues that understanding "ought" statements requires more than just logical implication; it necessitates causal knowledge to reason about the consequences of actions and how observations are established.
  • Qualitative Reasoning is Sufficient: By combining order-of-magnitude probabilities and utilities, the framework can make meaningful qualitative judgments about actions and consequences, avoiding the need for precise numerical values.
  • Persistence Model for Actions: A novel persistence model is introduced to handle the effects of actions. It allows for properties to change due to causal factors without necessarily implying an abnormality, distinguishing between causally supported changes and spontaneous ones.
  • Unified Framework: The approach integrates belief, utility, and causality within a single framework, allowing for a unified analysis of conditional "ought" statements and planning under uncertainty.

Key Results

  • Assertability Criterion for O(A|C): The paper provides a formal criterion (Eq. 14) for determining when a conditional "ought" statement is assertable, based on comparing the expected utility of doing A versus not doing A, given an observation C.
  • Post-Action Belief Update: A formula (Eq. 19) is derived for updating belief rankings after an action, incorporating causal influences and persistence assumptions.
  • Example Demonstrations: The paper illustrates the framework with examples, such as the "cloudy/umbrella" scenario and a dialogue between robots about a light switch, demonstrating how the framework can handle reversals of "ought" statements based on new observations.

Limitations and Future Work

  • Myopic Approximation: The current model assumes a myopic approximation for actions, where action A is either terminal or its consequences are fully embodied in the current belief and utility functions. This might not be suitable for complex sequential planning situations where actions are taken solely to enable future actions.
  • Encoding Complexity: While the framework is powerful, encoding the full epistemic state (κ, Γ, μ) can be complex, especially for large domains.
  • Future Work: The authors suggest exploring how their framework can be integrated with other formalisms, such as Boutilier's modal logic for conditional goals, and applying it to more complex decision-theoretic tasks like selecting optimal information-gathering strategies. The paper also hints at the need to explicitly address temporal aspects in more detail for sequential planning.

Summary updated

BibTeX
@inproceedings{pearl1993from,
  author = {Judea Pearl},
  title = {{From Conditional Oughts to Qualitative Decision Theory}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure