Reduced order modeling for stochastic prediction and data assimilation onboard autonomous platforms at sea

There are many significant challenges for unmanned autonomous platforms at sea including predicting the likely scenarios for the ocean environment, quantifying regional uncertainties, and updating forecasts of the evolving dynamics using their observations. Due to the operational constraints such as onboard power, memory, bandwidth, and space limitations, efficient adaptive reduced order models (ROMs) are needed for onboard predictions. In the first part, several reduced order modeling schemes for regional ocean forecasting onboard autonomous platforms at sea are described, investigated, and evaluated. We find that Dynamic Mode Decomposition (DMD), a data-driven dimensionality reduction algorithm, can be used for accurate predictions for short periods in ocean environments. We evaluate DMD methods for ocean PE simulations by comparing and testing several schemes including domain splitting, adjusting training size, and utilizing 3D inputs. Three new approaches that combine uncertainty with DMD are also investigated and found to produce practical and accurate results, especially if we employ either an ensemble of DMD forecasts or the DMD of an ensemble of forecasts. We also demonstrate some results from projecting / compressing high-fidelity forecasts using schemes such as POD projection and K-SVD for sparse representation due to showing promise for distributing forecasts efficiently to remote vehicles. In the second part, we combine DMD methods with the GMM-DO filter to produce DMD forecasts with Bayesian data assimilation that can quickly and efficiently be computed onboard an autonomous platform. We compare the accuracy of our results to traditional DMD forecasts and DMD with Ensemble Kalman Filter (EnKF) forecast results and show that in Root Mean Square Error (RMSE) sense as well as error field sense, that the DMD with GMM-DO errors are smaller and the errors grow slower in time than the other mentioned schemes. We also showcase the DMD of the ensemble method with GMM-DO. We conclude that due to its accurate and computationally efficient results, it could be readily applied onboard autonomous platforms. Overall, our contributions developed and integrated stochastic DMD forecasts and efficient Bayesian GMM-DO updates of the DMD state and parameters, learning from the limited gappy observation data sets.

Online Learning Probabilistic Event Calculus Theories in Answer Set Programming

Complex Event Recognition (CER) systems detect event occurrences in streaming time-stamped input using predefined event patterns. Logic-based approaches are of special interest in CER, since, via Statistical Relational AI, they combine uncertainty-resilient reasoning with time and change, with machine learning, thus alleviating the cost of manual event pattern authoring. We present a system based on Answer Set Programming (ASP), capable of probabilistic reasoning with complex event patterns in the form of weighted rules in the Event Calculus, whose structure and weights are learnt online. We compare our ASP-based implementation with a Markov Logic-based one and with a number of state-of-the-art batch learning algorithms on CER data sets for activity recognition, maritime surveillance and fleet management. Our results demonstrate the superiority of our novel approach, both in terms of efficiency and predictive performance. This paper is under consideration for publication in Theory and Practice of Logic Programming (TPLP).

WOLED: A tool for Online Learning Weighted Answer Set Rules for Temporal Reasoning Under Uncertainty

Complex Event Recognition (CER) systems detect event occurrences in streaming time-stamped input using predefined event patterns. Logic-based approaches are of special interest in CER, since, via Statistical Relational AI, they combine uncertainty-resilient reasoning with time and change, with machine learning, thus alleviating the cost of manual event pattern authoring. We present WOLED, a system based on Answer Set Programming (ASP), capable of probabilistic reasoning with complex event patterns in the form of weighted rules in the Event Calculus, whose structure and weights are learnt online. We compare our ASP-based implementation with a Markov Logic-based one and with a crisp version of the algorithm that learns unweighted rules, on CER datasets for activity recognition, maritime surveillance and fleet management. Our results demonstrate the superiority of our novel implementation, both in terms of efficiency and predictive performance.

Learning Interpretable Relational Structures of Hinge-loss Markov Random Fields

Statistical relational models such as Markov logic networks (MLNs) and hinge-loss Markov random fields (HL-MRFs) are specified using templated weighted first-order logic clauses, leading to the creation of complex, yet easy to encode models that effectively combine uncertainty and logic. Learning the structure of these models from data reduces the human effort of identifying the right structures. In this work, we present an asynchronous deep reinforcement learning algorithm to automatically learn HL-MRF clause structures. Our algorithm possesses the ability to learn semantically meaningful structures that appeal to human intuition and understanding, while simultaneously being able to learn structures from data, thus learning structures that have both the desirable qualities of interpretability and good prediction performance. The asynchronous nature of our algorithm further provides the ability to learn diverse structures via exploration, while remaining scalable. We demonstrate the ability of the models to learn semantically meaningful structures that also achieve better prediction performance when compared with a greedy search algorithm, a path-based algorithm, and manually defined clauses on two computational social science applications: i) modeling recovery in alcohol use disorder, and ii) detecting bullying.

Comparative Study of Rebound Hammer, Nitto Hammer, and Pullout Tests to Estimate Concrete In-Place Strength by Using Random Sampling Analysis

The use of nondestructive testing (NDT) techniques to estimate concrete in-place strength has been broadly studied, with proof of their usefulness in complementing destructive testing (DT). However, the use of DT techniques still dominates. The main objective of this investigation was to compare the performance of three NDT techniques—the rebound hammer, Nitto hammer, and pullout tests—to determine in-place strength. NDT-versus-strength correlation curves were fit to data measured from thick concrete slabs. Strength was measured from cast-in-place cylinders. Analyses of NDT sensitivity, uncertainty, and variability are presented. A new parameter to quantify the performance of the NDT techniques is proposed. This parameter is the limit error between the measured and estimated strengths, which combine uncertainty and variability analyses. The analysis shows that the least limit error for predicting in-place strength was achieved by the rebound hammer test when one testing location was considered or by the pullout test for two or more testing locations.