My lab investigates reliability measures to assess the performance of traffic systems subject to non-recurrent incidents through applied queueing theory. We study the density (congestion) properties of traffic when the system is subject to random deteriorations of the quality of service. For this, we consider stationary and analytical solutions for the number of customer’s distributions of queues with Markov-modulated service rates. The framework has the potential to contribute in several fields, including computer science (job scheduling, traffic of information on mobile networks), and transportation. We validate and calibrate these parametric distributions on a concrete example using data on roadway traffic travel time, traffic density, incident reports and weather conditions as sources of traffic deterioration. The measures defined for the performance of the system have explicit dependencies on traffic and incident parameters, thus avoiding the use of costly simulation. One of such measures is the probability of system breakdown, when the system has reached complete congestion. While current models for such measure are expensive and complicated, our model yields robust results with intuitive, easily estimated statistical parameters.

We use data to understand patterns of defect generation in rail systems to mitigate risks arising from such defects. The solution of this problem generalizes for several other risk-related fields. Currently, we look at how defect generation can be assessed under inspections via machine-learning algorithms. To account for the particular risk-aversion characteristic of this setting where false negatives are worse than false positives, we adapt the architectures and loss function to account for different levels of risk-aversion. These algorithms are scalable and transferable, thus yielding robust results. Current leading candidates include random forests and recurrent neural networks. Then, we develop algorithms based on game theory and dynamic programming that generate policies for optimally scheduling, thus providing thorough risk analysis and mitigation assessment. We show that the choice of nodes to inspect are indexable under a restless-multi-armed-bandit formulation, and derive the closed-form equation for calculating the indices. Our results are validated using data from a railway system, where defects occur randomly, and inspection crews are allocated to segments each period. Remaining work includes the development of formal analytical solutions to find equilibria on the game theory framework. Finally, we intend to study how the geography of the network plays a role in the scheduling decision. For this, given the intractability of the problem, we intend to use reinforcement learning methods as a means of finding a solution.

During my work at NASA, I implemented and developed AI models for the classification of star light-curves (containing a planet candidate or some other astrophysical entity). My team and I trained adaptations of preestablished CNN model architectures on large datasets (> 1 TB) and obtained better results than the benchmark ones (in accuracy, precision, and AUC). For such, we employed cluster computing via the super-computer at NASA, and we applied the BOHB (Bayesian Optimization and Hyperband) algorithm to determine the hyperparameters in the CNN. Furthermore, I utilized the trained model to infer whether the hundreds of thousands of stars that had been initially vetted by older systems indeed contained no planet. Results indicating a high likelihood of missed planets led me to code a large data-gathering algorithm to run follow-up reports for expert confirmations. This algorithm combines and processes data from various sources and of different formats, thus seamlessly generating the reports. It has been well-received by the data science group at NASA.