Publications

There have been some concerns that telecommunication companies are facing today which are spatial in nature. Systems operated by these telecommunication companies have evolved along the years. The main concern of these companies is the ability to provide quality service to customers or users in a dense regime. Therefore, they sought to answer questions such as (i) what is the best possible configurations of base stations and users that maximizes quality service? (ii) can one estimate and control the probability of bad service, which may be seen as a rare event? and many more which may arise during companies operations. To answer these questions one will often have to estimate the tail distribution of events, which may be considered under the realm of large deviations. In this article, we define the empirical marked measure of the Boolean model, which will serve estimate the intensity measure of the Marked Poisson Point Process of devices or users, and the empirical connectivity measure of the Boolean Model which will serve estimate the coverage probability density of the spatial telecommunication area. For these empirical measures, we establish large deviation principle (LDP) in the τ – topology.

Capturing significant events in detailed, high-fidelity HPC simulations for scientific investigations is becoming more and more challenging because exporting the entire simulation state at each timestep is often impractical. Important phases of events may be overlooked between checkpoints, and transient events might be entirely missed, complicating the detection of these events afterward. In this study, we concentrate on the in situ event detection use case of machine learning to optimize the amount of data saved to disk and the number of events captured. We present the In situ Machine Learning (ISML) framework which is composed from signature, measure and decision building blocks with well defined semantics. We show the capabilities of our framework using a synthetic data with well-known anomalies. We use manifold learning techniques such as Multidimensional Scaling (MDS), Isometric Mapping (Isomap), Locally Linear Embedding (LLE) and t-distributed Stochastic Neighbor Embedding (t-SNE) to create the signatures. Conveniently chosen measure and decision functions were used to detect the anomalies at various timesteps. The anomaly recall and the total percentage of flagged analysis partitions were then compared for the various methods.