David Koelle

Complexities and Challenges in the Use of Bayesian Belief Networks: Informing the Design of Causal Influence Models

Bayesian belief networks (BNs) are well-suited to capturing vague and uncertain knowledge. However, the capture of this knowledge and associated reasoning from human domain experts often requires specialized knowledge engineers responsible for translating the experts communications into BN-based models. Across application domains, we have analyzed how these models are constructed, refined, and validated with domain experts. From this analysis, we have identified key user-centered complexities and challenges that we have used to drive the selection of simplifying assumptions. This led us to develop computational techniques and user interface methods that leverage these same assumptions with the goal of improving the efficiency and ease with which expert knowledge can be expressed, verified, validated, and encoded. In this paper, we present the results of our analysis of BN construction, validation, and use. We discuss how these results motivated the design of a simplified version of BNs called Causal Influence Models (CIMs). In addition, we detail how CIMs enable the design and construction of user interface mechanisms that address complexities identified in our analysis.

A Toolkit to Assist Researchers to More Efficiently Conduct Experiments Assessing Human State

For teams to perform effectively, individuals must focus on their own tasks, while simultaneously maintaining awareness of other team members. Researchers studying and attempting to optimize performance of teams as well as individual team members use assessments of behavioral, neurophysiological, and physiological signals that correlate with individual and team performance. However, synchronizing data from multiple sensor devices can be difficult, and building and using models to assess human states of interest can be time-consuming and non-intuitive. To assist researchers, we built an Adaptable Toolkit for the Assessment and Augmentation of Performance by Teams in Real Time (ADAPTER), which provides a framework that flexibly integrates sensors and fuses sensor data to assess performance. ADAPTER flexibly integrates current and emerging sensors; assists researchers in creating and implementing models that support research on performance and the development of augmentation strategies; and enables comprehensive and holistic characterization of team member performance during real-time experimental protocols.

A Flexible Framework for the Creation of Narrative-Centered Tools

To better support the creation of narrative-centered tools, developers need a flexible framework to integrate, catalog, select, and reuse narrative models. Computational models of narrative enable the creation of software tools to aid narrative processing, analysis, and generation. Narrative-centered tools explicitly or implicitly embody one or more models of narrative by their definition. However, narrative model creation is often expensive and difficult with no guaranteed benefit to the end system. This paper describes our preliminary approach towards creating the SONNET narrative framework, a flexible framework to integrate, catalog, select, and reuse narrative models, thereby lowering development costs and improving benefits from each model. The framework includes a lightweight ontology language for the definition of key terms and interrelationships among them. The framework specifies model metadata to allow developers to discover and understand models more readily. We discuss the structure of this framework and ongoing development incorporating narrative models.