Defense Notices

Human computer interaction (HCI) and artificial intelligence (AI) research have greatly progressed over the years. Work in HCI aims to create cyberphysical systems that facilitate good interactions with humans, while artificial intelligence work aims to understand the causes of intelligent behavior and reproduce them on a computer. To this point, HCI systems typically avoid the AI problem, and AI researchers typically have focused on building system that work alone or with other AI systems, but de-emphasise human collaboration. In this thesis, we examine the role of episodic memory in constructing intelligent agents that can collaborate with and learn from humans. We present our work showing that agents with episodic memory capabilities can expose their internal decision-making process to users, and that an agent can learn relational planning operators from episodic traces.

This thesis documents the improvements made to a UHF radar system for snow accumulation measurements and the implementation of an airborne HF radar system for ice sounding. The HF sounder radar was designed to operate at two discrete frequency bands centered at 14.1 MHz and 31.5 MHz with a peak power level of 1 kW, representing an order-of-magnitude increase over earlier implementations. A custom transmit/receive module was developed with a set of lumped-element impedance matching networks suitable for integration on a Twin Otter Aircraft. The system was integrated and deployed to Greenland in 2016, showing improved detection capabilities for the ice/bottom interface in some areas of Jakobshavn Glacier and the potential for cross-track aperture formation to mitigate surface clutter. The performance of the UHF radar (also known as the CReSIS Accumulation radar) was significantly improved by transitioning from a single channel realization with 5-10 Watts peak transmit power into a multi-channel system with 1.6 kW. This was accomplished through developing custom transmit/receive modules capable of handling 400-W peak per channel and fast switching, incorporating a high-speed waveform generator and data acquisition system, and upgrading the baluns which feed the antenna elements. We demonstrated dramatically improved observation capabilities over the course of two different field seasons in Greenland onboard the NASA P-3.

The performance of several methods to estimate the number of source signals impinging on a sensor array are compared using a traditional simulator and their performance for synthetic aperture radar tomography is discussed as it is useful in the fields of radar and remote sensing when multichannel arrays are employed. All methods use the sum of the likelihood function with a penalty term. We consider two signal models for model selection and refer to these as suboptimal and optimal. The suboptimal model uses a simplified signal model and the model selection and direction of arrival estimation are done in separate steps. The optimal model uses the actual signal model and the model selection and direction of arrival estimation are done in the same step. In the literature, suboptimal model selection is used because of computational efficiency, but in our radar post processing we are less time constrained and we implement the optimal model for the estimation and compare the performance results. Interestingly we find several methods discussed in the literature do not work using optimal model selection, but can work if the optimal model selection is normalized. We also formulate a new penalty term, numerically tuned so that it gives optimal performance over a particular set of operating conditions, and compare this method as well. The primary contribution of this work is the development of an optimizer that finds a numerically tuned penalty term that outperforms current methods and discussion of the normalization techniques applied to optimal model selection. Simulation results show that the numerically tuned model selection criteria is optimal and that the typical methods do not do well for low snapshots which are common in radar and remote sensing applications. We apply the algorithms to data collected by the CReSIS radar depth sounder and discuss the results.

In addition to model order estimation, array model errors should be estimated to improve direction of arrival estimation. The implementation of a parametric-model is discussed for array calibration that estimates the first and second order array model errors. Simulation results for the gain, phase and location errors are discussed.

The Compact Muon Solenoid - Precision Proton Spectrometer (CMS-PPS) detector at the Large Hadron Collider (LHC) operates at high luminosity and is designed to measure forward scattered protons resulting from proton-proton interactions involving photon and Pomeron exchange processes. The PPS uses tracking and timing detectors for these measurements. The timing detectors measure the arrival time of the protons on each side of the interaction and their difference is used to reconstruct the vertex of the interaction. A good time precision (~10ps) on the arrival time is desired to have a good precision (~2mm) on the vertex position. The time precision is approximately equal to the ratio of the Root Mean Square (RMS) noise to the slew rate of the signal obtained from the detector.

Components of the timing detector include Ultra-Fast Silicon Detector (UFSD) sensors that generate a current pulse, transimpedance amplifier with shaping, and a discriminator. This thesis discusses the circuit schematic and simulations of an amplifier designed to have a time precision and the choice and simulation of discriminators with Low Voltage Differential Signal (LVDS) outputs. Additionally, details on the Printed Circuit Board (PCB) design including arrangement of components, traces, and stackup have been discussed for a 6-layer PCB that houses these three components. The PCB board has been manufactured and test results were performed to assess the functionality.

The Internet of Things (IoT) is evolving rapidly to every aspect of human life including healthcare, homes, cities, and driverless vehicles that makes humans more dependent on the Internet and related infrastructure. While many researchers have studied the structure of the Internet that is resilient as a whole, new studies are required to investigate the resilience of the edge networks in which people and “things” connect to the Internet. Since the range of service requirements varies at the edge of the network, a wide variety of protocols are needed. In this research proposal, we survey standard protocols and IoT models. Next, we propose an abstract model for smart homes and cities to illustrate the heterogeneity and complexity of network structure. Our initial results show that the heterogeneity of the protocols has a direct effect on the IoT and smart city resilience. As the next step, we make a graph model from the proposed model and do graph theoretic analysis to recognize the fundamental behavior of the network to improve its robustness. We perform the process of improvement through modifying topology, adding extra nodes, and links when necessary. Finally, we will conduct various simulation studies on the model to validate its resilience.

People use online social networks like Twitter to communicate and discuss a variety of topics. This makes these social platforms an import source of information. In the case of Twitter, to make sense of this source of information, understanding the content of tweets is important in understanding what is being discussed on these social platforms and how ideas and opinions of a group are coalescing around certain themes. Although there are many algorithms to classify(identify) the topics, the restricted length of the tweets and usage of jargon, abbreviations and urls make it hard to perform without immense expertise in natural language processing. To address the need for content analysis in twitter that is easily implementable, we introduce two measures based on the term frequency to identify the topics in the twitter microblogging environment. We apply these measures to the tweets with hashtags related to the Pulse night club shooting in Orlando that happened on June 12, 2016. This event is branded as both terrorist attack and hate crime and different people on twitter tweeted about this event differently forming social network communities, making this a fitting domain to explore our algorithms ability to detect the topics of community discussions on twitter.  Using community detection algorithms, we discover communities in twitter. We then use frequency statistics and Monte Carlo simulation to determine the significance of certain hashtags. We show that this approach is capable of uncovering differences in community discussions and propose this method as a means to do community based content detection.

On June 12, 2016, 49 people were killed and another 58 wounded in an attack at Pulse Nightclub in Orlando Florida. This incident was regarded as both hate crime against LGBT people and as a terrorist attack. This project focuses on analyzing tweets a week after the terrorist attack, specifically looking at how different communities within twitter were discussing this event. To understand how the twitter users in different communities are discussing this event, a set of hashtag frequency-based evaluation measures and simulations are proposed. The simulations are used to assess the specific hashtag content of a community. Using community detection algorithms and text analysis tools, significant topics that specific communities are discussing and  the topics that are being avoided by those communities are discovered.

Rule Induction is one of the major applications of rough set theory. However, traditional rough set models cannot deal with incomplete data sets. Missing values can be handled by data pre-processing or extension of rough set models. Two data pre-processing methods and one extension of the rough set model are considered in this project. These being filling in missing values with most common data, ignoring objects by deleting records and extended discernibility matrix. The objective is to compare these methods in terms of stability and effectiveness. All three methods have same rule induction method and are analyzed based on test accuracy and missing attribute level percentage. To better understand the properties of these approaches, eight real-life data-sets with varying level of missing attribute values are used for testing. Based on the results, we discuss the relative merits of three approaches in an attempt to decide upon optimal approach. The final conclusion is that the best method is to use a pre-processing method which is filling in missing values with most common data.​

Recently machine learning has created significant advancements in many areas like health, finance, education, sports, etc. which has encouraged the development of many predictive models. In machine learning, we extract hidden, previously unknown, and potentially useful high-level information from low-level data. Cross-validation is a typical strategy for estimating the performance. It simulates the process of fitting to different datasets and seeing how different predictions can be. In this project, we review accuracy estimation methods and compare two resampling methods, such as leaving-one-out and bootstrap. We compared these validation methods using LEM1 rule induction algorithm. Our results indicate that for real-world datasets similar to ours, bootstrap may be optimistic.

With the advancement of arbitrary waveform generation techniques, new radar transmission modes can be designed via precise control of the waveform's time-domain signal structure. The finer degree of emission control for a waveform (or multiple waveforms via a digital array) presents an opportunity to reduce ambiguities in the estimation of parameters within the radar backscatter. While this freedom opens the door to new emission capabilities, one must still consider the practical attributes for radar waveform design. Constraints such as constant amplitude (to maintain sufficient power efficiency) and continuous phase (for spectral containment) are still considered prerequisites for high-powered radar waveforms. These criteria are also applicable to the design of multiple waveforms emitted from an antenna array in a multiple-input multiple-output (MIMO) mode.

In this work, two spatially-diverse radar emission design methods are introduced that provide constant amplitude, spectrally-contained waveforms. The first design method, denoted as spatial modulation, designs the radar waveforms via a polyphase-coded frequency-modulated (PCFM) framework to steer the coherent mainbeam of the emission within a pulse. The second design method is an iterative scheme to generate waveforms that achieve a desired wideband and/or widebeam radar emission. However, a wideband and widebeam emission can place a portion of the emitted energy into what is known as the `invisible' space of the array, which is related to the storage of reactive power that can damage a radar transmitter. The proposed design method purposefully avoids this space and a quantity denoted as the Fractional Reactive Power (FRP) is defined to assess the quality of the result.

The design of FM waveforms via traditional gradient-based optimization methods is also considered. A waveform model is proposed that is a generalization of the PCFM implementation, denoted as coded-FM (CFM), which defines the phase of the waveform via a summation of weighted, predefined basis functions. Therefore, gradient-based methods can be used to minimize a given cost function with respect to a finite set of optimizable parameters. A generalized integrated sidelobe metric is used as the optimization cost function to minimize the correlation range sidelobes of the radar waveform