Probabilistic Temporal Reasoning with Endogenous Change

Steve Hanks, David Madigan, Jonathan GavrinUAI 1995

TL;DR

This paper introduces a novel probabilistic model for reasoning about dynamic systems that change over time due to both external events and internal processes. It addresses the challenge of modeling complex medical scenarios where diagnosis and treatment decisions are influenced by the evolving state of the patient, proposing a method for eliciting and solving such models using sequential imputation.

What Problem Does This Paper Address?

The paper tackles the problem of "temporal decision problems," particularly in domains like medical diagnosis and treatment. These are situations where the time it takes to gather information or decide on a course of action is comparable to the time it takes for the patient's condition to change significantly. This means that both external factors (like accidents, treatments, or tests) and internal factors (like disease progression) must be considered. Traditional AI temporal reasoning often struggles with explicitly modeling these endogenous changes and the uncertainty associated with them, especially when trying to elicit this knowledge from human experts. The authors aim to bridge this gap by providing a framework that can capture both exogenous and endogenous influences in a probabilistic manner, making it suitable for expert elicitation and computational reasoning.

Methods

The core of the paper is a probabilistic model that represents the system's state using a vector of attributes, each with possible values. The model explicitly distinguishes between:

  • Exogenous Events: These are external occurrences (collisions, treatments, observations) that instantaneously change the system's state. The paper models these as probabilistic transformations from one state to another, represented by sets of consequences with associated probabilities. This approach is designed to be more compact than traditional Markov transition matrices by only specifying relevant attributes and changes.
  • Endogenous Change: This refers to changes within the system due to internal processes (e.g., brain swelling, internal bleeding). The model captures these using expert-elicited rules that describe how certain states (like head injury) lead to subsequent changes in other attributes (like pupil dilation) over specific time intervals. These rules are then combined using simple linear models to aggregate influences and estimate transition times.
  • Inference: To solve the model and make predictions or diagnoses, the authors propose a simulation-based approach using sequential imputation. This is an importance sampling technique that allows for incremental updates to the model as new observations become available, making it efficient and adaptable to extending the scenario.

Datasets

Not discussed in the paper. The paper uses a simplified example of vehicle accident trauma for illustration and evaluation.

Key Insights and Ideas

  • Eliciting Dynamics is Natural: The authors found that human experts are more comfortable reasoning about the evolution of a system over time (diachronic relationships) than about static relationships at a single point in time (synchronic constraints). This insight drives their approach of eliciting dynamic models directly.
  • Bridging AI and Statistics: The paper aims to synthesize work from both AI (rule-based temporal reasoning) and statistics (probabilistic modeling), addressing the common weakness of AI in handling endogenous change and uncertainty.
  • Sequential Imputation for Adaptability: The use of sequential imputation is a key methodological contribution, allowing the model to be updated incrementally with new observations without discarding previous computations. This is crucial for real-time decision-making in dynamic environments.
  • Modeling Uncertainty in Transition Times: The approach of using intervals for transition times and combining them linearly provides a practical way to handle expert uncertainty about the speed of endogenous changes.

Key Results

The paper demonstrates its approach on a simplified vehicle accident trauma scenario. They show that their sequential imputation method can produce reasonable accuracy in estimating probabilities (e.g., for collision severity and internal bleeding) conditioned on observations, even with a small number of trials. The results in Table 1 indicate that as the number of trials increases, the imputed probabilities converge towards the "exact" probabilities, and the "effective sample size" (ESS) suggests the efficiency of the sampling process. The method is shown to be compatible with real-time decision-making constraints.

Limitations and Future Work

The authors acknowledge several limitations:

  • Arbitrary Influence Combination: The linear models used to combine concordant and contrary influences are described as simple and arbitrary. While they seem to align with expert intuition, more sophisticated methods might be needed for other domains.
  • Elicitation of Endogenous Models: The techniques for eliciting endogenous models are still being developed, and the authors hope to create a variety of methods adaptable to different domains.
  • Computational Efficiency: While sequential imputation is efficient, substantial work is still needed to determine reasonable sample sizes for particular problems and to optimize the code.

Future work includes exploring more sophisticated methods for combining influences, developing a broader range of endogenous model elicitation techniques, and investigating hybrid models that combine diachronic and synchronic constraints.

Summary updated

BibTeX
@inproceedings{hanks1995probabilistic,
  author = {Steve Hanks and David Madigan and Jonathan Gavrin},
  title = {{Probabilistic Temporal Reasoning with Endogenous Change}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1995},
}
Figure