Christie Nelson

Nuclear detection using higher order learning

The detection of potentially threatening nuclear materials is a challenging homeland security problem. This research reports on the application of a novel statistical relational learning algorithm, Higher Order Naïve Bayes (HONB), to improve the detection and identification of nuclear isotopes. When classifying nuclear detection data, distinguishing potentially threatening from harmless radioisotopes is critical. These also must be distinguished from the naturally occurring radioactive background. This research applied Higher Order Learning to nuclear detection data to improve the detection and identification of four isotopes: Ga67, I131, In111, and Tc99m. In the research traditional IID machine learning methods are applied to the area of nuclear detection, and the results compared with the performance of leveraging higher-order dependencies between feature values using HONB. The findings give insight about the performance of higher-order classifiers (described in [2]) on datasets with small numbers of positive instances. In the initial study, Naïve Bayes was compared with its higher-order counterpart, Higher Order Naïve Bayes. HONB was found to perform statistically significantly better for isotope Ga67 when using a preprocessing methodology of discretizing then binarizing the input sensor data. Similar results were seen for different amounts of training data for I131, In111, and Tc99m. HONB was also found to perform statistically significantly better for isotopes I131 and Tc99m when the preprocessing involved normalization, discretization then binarization. This study shows that Higher Order Learning techniques can be very useful in the arena of nuclear detection.

Experimental designs for testing metal detectors at a large sports stadium

When utilizing metal detectors at a large venue such as a sports stadium, there are the competing objectives of accuracy of the patron screening and the speed of throughput. This research, carried out in collaboration with the security staff at MetLife Stadium in New Jersey as well as other stadiums, analyzed two patron screening methods: handheld metal detectors (“wands”) and walk-through metal detectors (“walk-throughs”). An initial experimental design was created to understand the effectiveness of wanding. This design was used with MetLife Stadium security during three training sessions. The data collected was used to understand (a) if the prohibited item was found and (b) in how many attempts. Various prohibited as well as allowable metal items were hidden at random in various locations on the body of individuals, who were then scored based on the importance weight of the item (guns were given more weight than keys for example). Trainees were then assigned a performance score based on speed and accuracy and were tested until they reached a minimum required score. Building on this initial experiment, a second more formal experiment was created to help MetLife Stadium staff understand how walk-throughs would perform outdoors at different security settings. This experiment focused less on training security staff and more on understanding the performance of walk-throughs in real situations (as opposed to idealized lab situations). This experiment was created to understand the walk-through performance at each setting in the outdoor environment; e.g., does a walk-through catch each of the pre-specified prohibited items, and is this consistent across machines on the same setting? Because of the number of factors to be considered (type of item, location, orientation, walk-through setting, etc.), designing the experiment required a sophisticated approach called Combinatorial Experimental Design. The experiment was part of two DHS-supported projects on best practices for stadium security.

Applying Topic Modeling to Railroad Grade Crossing Accident Report Text

The FRA railroad grade crossing accident database contains text comment fields that may provide additional information about grade crossing accidents. New text mining algorithms provide the potential to automatically extract information from text that can enhance traditional numeric analyses. Topic modeling algorithms are statistical methods that analyze the words of original texts to automatically discover the themes that run through them. A frequently used topic-modeling algorithm is Latent Dirichlet Analysis (LDA). In this paper we will show several examples of how labeled LDA can be applied to the FRA grade crossing data to better understand categories of words and phrases that are associated with various types of grade crossing accidents.© 2015 ASME

Walk-through metal detectors for stadium security

Walk-through metal detectors (WTMDs) are increasingly utilized as a security measure at large events held at stadiums. As this is a different environment than standard use cases (e.g. airports, prisons), work has been done to understand how to best integrate this technology into an outdoor high throughput screening environment. We have performed several experiments on these WTMDs to understand their performance in real stadium environments in order to inform security directors of areas of potential vulnerabilities. These experiments also help the security directors to be more informed of the real functionality in the field of the WTMDs, as they were examined to understand performance in non-laboratory settings. Experimental design was utilized to create three types of experiments which were performed on three different brands of WTMDs. These experiments focused on height and orientation, speed passing through the WTMD, and proximity outside of the WTMD. These experiments will help security personnel at large stadium venues gain a greater understanding of how WTMDs work and how they perform.

The ACCAM model: simulating aviation mission readiness for U.S. coast guard stations

We present a model and discrete event simulation of USCG Air Stations, accounting for the mission demands and maintenance procedures pertaining to USCG aircraft. The simulation provides aircraft availability distributions and mission performance metrics based on varying input scenarios, including changes in the number of stationed aircraft and maintenance targets. The Air Station model is novel in its relatively simple, easily tunable, renewal process treatment of maintenance procedures, mitigating the need for the modeling of complex maintenance subprocesses and the resulting statistical estimation of numerous parameters. The simulation also models mission requirements such as Search and Rescue that are stochastic in time and space. Simulations are consistent with historical data and offer insights into hypothetical scenarios.

To be or not to be IID: Can Zipf's Law help?

Classification is a popular problem within machine learning, and increasing the effectiveness of classification algorithms has many significant applications within industry and academia. In particular, focus will be given to Higher-Order Naive Bayes (HONB), a relational variant of the famed Naive Bayes (NB) statistical classification algorithm that has been shown to outperform Naive Bayes in many cases [1,10]. Specifically, HONB has outperformed NB on character n-gram based feature spaces when the available training data is small [2]. In this paper, a correlation is hypothesized between the performance of HONB on character n-gram feature spaces and how closely the feature space distribution follows Zipfs Law. This hypothesis stems from the overarching goal of ultimately understanding HONB and knowing when it will outperform NB. Textual datasets ranging from several thousand instances to nearly 20,000 instances, some containing microtext, were used to generate character n-gram feature spaces. HONB and NB were both used to model these datasets, using varying character n-gram sizes (2-7) and dictionary sizes up to 5000 features. The performances of HONB and NB were then compared, and the results show potential support for our hypothesis: namely, the results support the hypothesized correlation for the Accuracy and Precision metrics. Additionally, a solution is provided for an open problem which was presented in [1], giving an explicit formula for the number of SDRs from k given sets, which has connections to counting higher-order paths of arbitrary length, which are important in the learning stage of HONB.

Optimization of emergency response using higher order learning and clustering of 911 text messages

In real-time emergency response an accurate picture of the situation is needed quickly. Often during large-scale disasters, cell towers become overloaded, and the only way of communication is through text messages. It becomes important to gather information from text messages sent to emergency numbers in order to respond quickly and efficiently with life-saving efforts. In addition, responders are unable to manually handle the large volume of incoming texts. To add to this difficult problem, these data sources tend to be microtext. This research developed a methodology to summarize text messages sent during an emergency, including analysis of locations. The real-time disaster needs were then input into a mixed integer programming resource allocation model for distribution of resources for disaster aid. Prior research included resource allocation and text modeling, but the combination of the two is a novel application not only in this arena, but more broadly across domains.