Process, structure and modularity in reasoning with uncertainty

Bruce D'AmbrosioUAI 1988

TL;DR

This paper introduces Hybrid Uncertainty Management (HUM), a novel framework for AI systems that goes beyond simple probabilistic inference. HUM focuses on the process of reasoning with uncertainty, enabling interactive model building, revision, and defeasible reasoning, while still adhering to normative probabilistic models when data is complete. It addresses the limitations of existing systems by providing a structured, symbolic representation that supports both the "micro-management" of individual uncertain events and the "macro-management" of how an agent constructs and modifies its understanding of the world.

What Problem Does This Paper Address?

Current AI systems for reasoning under uncertainty often focus heavily on the "micro-management" of uncertainty – calculating probabilities for specific events. However, this overlooks the crucial "macro-management" aspects: how an intelligent agent interactively formulates problems, incrementally builds and revises its models of the world, and makes decisions based on uncertain information. Existing systems struggle with the dynamic nature of real-world problems, where knowledge is often incomplete, and the structure of evidential relationships can change. This paper aims to bridge this gap by providing a computational approach that supports the entire process of uncertainty management, from initial problem formulation to dynamic model revision, while maintaining a connection to normative probabilistic reasoning.

Methods

The core of HUM is built upon an Assumption-Based Truth Maintenance System (ATMS). This provides a robust foundation for propositional logic and allows for the representation of uncertainty in a structured, symbolic way. Key methodological contributions include:

  • Explicit, Local Representations: Information, including uncertainty, is represented locally and explicitly, preserving both the content and structure of the information.
  • Incremental Construction and Evaluation: HUM supports the dynamic construction, modification, and evaluation of belief networks. This means the system can reason about the model itself, not just the domain it represents.
  • Hybrid Approach: HUM combines the strengths of ATMS for symbolic reasoning and structured representation with mechanisms for probabilistic inference.
  • Micro-management: This involves interpreting ATMS assumptions as markers for probability distributions and developing mappings for basic probabilistic inference computations. The ATMS "justification" is used to represent probabilistic relationships, and its environment propagation mechanism handles inference.
  • Macro-management: This focuses on reasoning about model construction decisions and revising domain models when inference reveals problems. ATMS assumptions are used to record defeasible model structure decisions, allowing for dynamic extension and revision. A mixed forward and backward chaining rule language is used for model construction.

Datasets

Not discussed in the paper. The paper uses illustrative examples derived from existing literature on uncertainty in AI (e.g., the "Three Urns" problem and the "Chernobyl" example).

Key Insights and Ideas

  • Process is as Important as Probability: The paper argues that a complete approach to uncertainty management must address the process of reasoning, not just the calculation of probabilities. This includes interactive problem formulation, incremental model building, and defeasible decision-making.
  • Structure Supports Modularity: By representing uncertainty symbolically and structurally, HUM addresses concerns about modularity. Rules and relationships can be expressed once and remain correct for various scenarios, avoiding the need to re-express them for each new piece of evidence.
  • Defeasible Reasoning for Model Structure: ATMS assumptions are used to represent defeasible decisions about model structure. This allows for the dynamic extension and revision of domain models, crucial for adapting to new information.
  • Bridging Symbolic and Numeric Reasoning: HUM effectively integrates symbolic reasoning capabilities of ATMS with probabilistic inference, allowing for reasoning with incomplete information while conforming to normative models when data is available.
  • Incrementalism is Key: The ability to incrementally extend and update belief networks and their representations is highlighted as an essential aspect of uncertainty reasoning that has been overlooked.

Key Results

The paper demonstrates HUM's capabilities through two detailed examples:

  • Three Urns Example: This example shows how HUM can incrementally update posterior probabilities for hypotheses (e.g., which urn is being used) as new evidence (bean draws) is observed. It illustrates bidirectional inference and the ability to represent classes of variables.
  • Thousands Dead Example (Chernobyl): This example showcases HUM's ability to handle correlated evidence and, more importantly, to reason about and revise the structure of evidential relationships. It demonstrates how the system can retract assumptions about independence when new information suggests otherwise, leading to a restructured model and updated beliefs. The paper emphasizes that this structural revision is more efficient than rebuilding the entire model.

Limitations and Future Work

  • Control of Reasoning: The current mechanisms for controlling the reasoning process are described as crude and only handle a few special cases. The authors plan to incorporate their approach into a blackboard-like architecture for more flexibility.
  • Decision Analysis: The machinery of decision analysis has barely begun to be incorporated into the framework.
  • Explanation: The structured symbolic representation provides a fertile ground for exploring the problem of explanation of results, a topic not yet explored.
  • Performance: The current implementation is an early prototype, and its speed is only marginally adequate for the experiments the authors wish to run.
  • Other Formalisms: While the paper focuses on probabilistic relationships, the authors are developing mappings for other uncertainty formalisms like belief functions and fuzzy set theory to understand their impact on reasoning.

Summary updated

BibTeX
@inproceedings{dambrosio1988process,
  author = {Bruce D'Ambrosio},
  title = {{Process, structure and modularity in reasoning with uncertainty}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1988},
}
Figure