Probabilistic Learning of Operator Interest in Surveillance Environments for Online Track Characterization

Kravitz, Eli

Graduate Thesis Or Dissertation

Probabilistic Learning of Operator Interest in Surveillance Environments for Online Track Characterization Public Deposited

Analytics

Download PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/wm117q65r

Abstract

Understanding user interests and behaviors is an important aspect to consider when developing machine learning models to augment data-intensive human workflows. By understanding what a user desires, these models can help prioritize workflow and essential information for decision making the way a human would, and form the baseline for trusted autonomous systems. This work considers the problem of human interest classification in the context of missile defense surveillance. In this case, users are satellite operators who must prioritize simultaneous processing and characterization of multiple (e.g. hundreds of) target tracks across the globe for extended periods of time under strict time and accuracy constraints.

The given problem is particularly challenging because user interest is not generally static, tracks have a short lifespan, and the user pool is small. Here, the problem is formulated similarly to a binary Bayesian logistic regression problem to classify operator interest in a given candidate track, but with the added complexity of partially observable input track feature variables and dependencies between these variables. Prior feature weight distributions are developed based on human inputs, and labeled training data is generated from online user interactions with the system. Derivations for a novel probabilistic model lead to intractable conditional posterior distributions for regression feature weights. These distributions are approximated in real-time online. Given trained weight distributions, human interest for unlabeled candidate tracks is inferred.

The developed interest classification model is validated using a simulated truth model by examining average track interest, learning rate, and Discounted Cumulative Gain for assessing relevancy of classified tracks. Validation is performed on several use cases, including an ideal case as well as several non-ideal cases. The results indicate that the model is robust to challenging cases, and is able to accurately learn features that interest a human operator. The introduction of a probabilistic model allowed for interest to be learned with limited prior information or labeled training data, and addressed the challenging problem of inferring user interest in a surveillance environment.

Creator