Master of Science in Engineering, University of Toledo, 2013, Engineering (Computer Science)

Radon-222 and its parent Radium-226 are naturally occurring radioactive decay products of Uranium-238. The US Environmental Protection Agency (USEPA) attributes about 10 percent of lung cancer cases that is `around 21,000 deaths per year' in the United States, caused due to indoor radon. The USEPA has categorized Ohio as a Zone 1 state (i.e. the average indoor radon screening level greater than 4 picocuries per liter). In order to implement preventive measures, it is necessary to know radon concentration levels in all the zip codes of a geographic area. However, it is not possible to survey all the zip codes, owing to reasons such as inapproachability. In such places where radon data are unavailable, several interpolation techniques are used to estimate the radon concentrations. This thesis presents a comparison between recently developed interpolation techniques to new techniques such as Support Vector Regression (SVR), and Random Forest Regression (RFR). Recently developed interpolation techniques include Artificial Neural Network (ANN), Knowledge Based Neural Networks (KBNN), Correction-Based Artificial Neural Networks (CBNN) and the conventional interpolation techniques such as Kriging, Local Polynomial Interpolation (LPI), Global Polynomial Interpolation (GPI) and Radial Basis Function (RBF) using the K-fold cross validation method.

Committee: William Acosta (Committee Chair); Vijay Devabhaktuni (Committee Co-Chair); Ashok Kumar (Committee Member); Rob Green (Committee Member) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2019, Geodetic Science

The study of harmful algal blooms is a well-established field with several conventional approaches such as processing samples, field observations, and examining weather and environmental effects that could change how the algae develop. However, with the frequency and severity of these blooms increasing worldwide, it has become progressively tougher to properly and efficiently assess their development and spread with the limited resources available. Therefore, assessment of the severity of algal blooms through remote sensing is helpful, especially since data from satellite imagery is already widely available for public use. Subjective judgments of existing data could yield additional human expertise that could be used in conjunction with the existing biology of bloom development to provide a more efficient framework to base future studies from. What is introduced in this dissertation is the development of a deductive technique based on a combination of fuzzy fault tree analysis and artificial neural network image recognition, utilizing readily available data from already well-established remote sensing equipment. The main advantage of this is that it could provide results that are just as accurate and yet could accept far more flexible inputs not as dependent on strict boundaries regarding lake conditions to compensate for limitations in gathering a priori knowledge. At the same time, the process is highly customizable and user-oriented, so that an assessor utilizing the interface need not necessarily have a comprehensive understanding of the underlying logic to interpret the results accordingly. Testing the interface and procedure on actual data arising from in situ sampling and satellite imagery through the fuzzy fault tree and convolutional neural network proved to give rise to accurate and logically consistent results, with the assessment of conditions arising from the fault tree consistent with the initial factors; the neural network also achieved a high d (open full item for complete abstract)

Committee: Tarunjit Butalia (Committee Chair); Michael Durand (Advisor); Steven Lower (Committee Member) Subjects: Earth; Geographic Information Science; Logic; Public Health

Master of Sciences (Engineering), Case Western Reserve University, 2017, EECS - Computer and Information Sciences

This work investigates a strategy for evaluating the navigability of terrain from 3-D imaging. Labeled training data was automatically generated by running a simulation of a mobile robot nai¨vely exploring a virtual world. During this exploration, sections of terrain were perceived through simulated depth imaging and saved with labels of safe or unsafe, depending on the outcome of the robot's experience driving through the perceived regions. This labeled data was used to train a deep convolutional neural network. Once trained, the network was able to evaluate the safety of perceived regions. The trained network was shown to be effective in achieving safe, autonomous driving through novel, challenging, unmapped terrain.

Committee: Wyatt Newman (Advisor); Cenk Cavusoglu (Committee Member); Michael Lewicki (Committee Member) Subjects: Computer Science; Robotics; Robots

Master of Science in Engineering, University of Akron, 2010, Biomedical Engineering

There has been a significant growth in the application of Artificial Neural Networks (ANNs) in the medical field including clinical decision support systems which call for high reliability and performance often involving multi-classification problems. Committee Neural Networks (CNNs) were developed for increased reliability and performance and were successfully applied to several multi-classification problems. However, the effect of the number of input parameters on the performance of the CNNs has not been investigated. The purpose of the present study was to investigate this effect on the CNN performance in a multi-classification problem of facial image based mood detection. Kulkarni et al (2009) used 15 parameters to develop a CNN system for classifying a given facial image into one of the six basic moods. Six subsets of parameters, for each image, were extracted from the parameter database used by Kulkarni et al, with an increasing number of parameters (from five to ten). The entire data was divided into training data, initial testing data, and final evaluation data. The training data from the six subsets were used to train six groups of neural networks and these networks were subjected to initial testing using the corresponding initial testing data. CNNs of different sizes were formed for each group by selecting the best performing networks based on the initial testing results. These CNNs were further tested using the initial testing data and the best performing CNN from each group was selected for further evaluation. All the selected committees were further evaluated using the final testing data from subjects not used in training or initial testing. Two approaches were used for converting neural network output into binary: “winner takes all” approach and the “thresholding” approach. The results from both the approaches show that with an increasing number of input parameters, the accuracy increased initially and then decreased. The committee decision was more ac (open full item for complete abstract)

Committee: Dr. Narender Reddy Dr. (Advisor) Subjects: Engineering

Master of Science, University of Akron, 2008, Computer Science

Internet technology has made international communication easy and convenient. This convenience has compelled a number of people to rely on electronic mail for almost all spheres of life – personal, business etc. Scrupulous organizations/individuals have taken undue advantage of this convenience and populate users' inboxes with unwanted messages making email spam a menace. Even as anti-spam software producers think they have almost solved the problem, spammers come out with new techniques. One such tactic in the spammers' toolbox comes in the form of image spam – messages that contain little more than a link to an image rendered in an HTML mail reader. The image typically contains the spam message one hopes to avoid, yet it is able to bypass most filters due to the composition and format of these pictures. This research focuses on identifying these images as spam by using an artificial neural network (ANN), software programs used for recognizing patterns, based on the biological neural networks in our brains. As information propagates through a neural network, it “learns” about the data. A large collection of both spam and non-spam images have being used to train an ANN, and then test the effectiveness of the trained network against an unidentified or already identified set of pictures. This process involves formatting images and adding the desired training values expected by the ANN. Several different ANNS have being trained using different configurations of hidden layers and nodes per layer. A detailed process for preprocessing spam image files is given, followed by a description on how to train an artificial neural network to distinguish between ham and spam. Finally, the trained network is tested against both known and unknown images.

Committee: Kathy Liszka PhD (Advisor); Timothy O’Neil (Other); Tim Marguish (Other) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2021, Physics

Reservoir computing (RC) is a machine learning method especially well suited to solving physical problems, by using an internal dynamic system known as a 'reservoir'. Many systems are suitable for use as an internal reservoir. A common choice is an echo state network (ESN), a network with recurrent connections that gives the RC a memory which it uses to efficiently solve many time-domain problems such as forecasting chaotic systems and hidden state inference. However, constructing an ESN involves a large number of poorly- understood meta-parameters, and the properties that an ESN must have to solve these tasks well are largely unknown. In this dissertation, I explore what parts of an RC are absolutely necessary. I build ESNs that perform well at system forecasting despite an extremely simple internal network structure, without any recurrent connections at all, breaking one of the most common rules of ESN design. These simple reservoirs indicate that the role of the reservoir in the RC is only to remember a finite number of time-delays of the RCs input, and while a complicated network can achieve this, in many cases a simple one achieves this as well. I then build upon a recent proof of the equivalence between a specific ESN construction and the nonlinear vector auto-regression (NVAR) method with my collaborators. The NVAR is an RC boiled down to its most essential components, taking the necessary time- delay taps directly rather than relying on an internal dynamic reservoir. I demonstrate these RCs-without-reservoirs on a variety of classical RC problems, showing that in many cases an NVAR will perform as well or better than an RC despite the simpler method. I then conclude with an example problem that highlights a remaining unsolved issue in the application of NVARs, and then look to a possible future where NVARs may supplant RCs.

Committee: Daniel Gauthier (Advisor); Amy Connolly (Committee Member); Ciriyam Jayaprakash (Committee Member); Gregory Lafyatis (Committee Member) Subjects: Physics

Doctor of Philosophy, The Ohio State University, 2019, Physics

There is currently great interest in applying artificial neural networks to a host of commercial and industrial tasks. Such networks with a layered, feedforward structure are currently deployed in technologies ranging from facial recognition software to self-driving cars. They are favored by a large portion of machine learning experts for a number of reasons. Namely: they possess a documented ability to generalize to unseen data and handle large data sets; there exists a number of well-understood training algorithms and integrated software packages for implementing them; and they have rigorously proven expressive power making them capable of approximating any bounded, static map arbitrarily well. Within the last couple of decades, reservoir computing has emerged as a method for training a different type of artificial neural network known as a recurrent neural network. Unlike layered, feedforward neural networks, recurrent neural networks are non-trivial dynamical systems that exhibit time-dependence and dynamical memory. In addition to being more biologically plausible, they more naturally handle time-dependent tasks such as predicting the load on an electrical grid or efficiently controlling a complicated industrial process. Fully-trained recurrent neural networks have high expressive power and are capable of emulating broad classes of dynamical systems. However, despite many recent insights, reservoir computing remains relatively young as a field. It remains unclear what fundamental properties yield a well-performing reservoir computer. In practice, this results in their design being left to domain experts, despite the actual training process being remarkably simple to implement. In this thesis, I describe a number of numerical and experimental results that expand the understanding and application of reservoir computing techniques. I develop an algorithm for controlling unknown dynamical systems with layers of reservoir computers. I demonstrate this algori (open full item for complete abstract)

Committee: Daniel Gauthier (Advisor); Gregory Lafyatis (Committee Member); Dick Furnstahl (Committee Member); Mikhail Belkin (Committee Member); Christopher Zirkle (Committee Member) Subjects: Physics

PhD, University of Cincinnati, 2012, Engineering and Applied Science: Computer Science and Engineering

The classical approach to artificial intelligence (i.e. symbol manipulation) and the connectionist approach (artificial neural networks) have been criticized for their inadequacies. The philosopher John Searle's Chinese room thought experiment argued that symbolic systems have no understanding of the meaning contained in their representations. The philosophers Jerry Fodor and Zenon Pylyshyn argued that artificial neural networks could not exhibit certain features of human cognition, such as systematicity and composition of representations. We take the view that both of these problems can be solved by a suitable integration of connectionist and symbolic systems. In this work we investigate methods of using artificial neural networks to produce descriptions in propositional and predicate logic. Artificial neural networks are stuctured such that, upon training, simple features of the network correspond directly to either propositional variables in one case, and objects and predicates in the other. In both cases, the methods were tested on character recognition tasks.

Committee: Raj Bhatnagar PhD (Committee Chair); Yizong Cheng PhD (Committee Member); Carla Purdy PhD (Committee Member); George Purdy PhD (Committee Member); John Schlipf PhD (Committee Member) Subjects: Computer Science

Doctor of Philosophy in Engineering, University of Toledo, 2012, Civil Engineering

The objective of this research was to develop and evaluate an integrated approach to model the occupant exposure to in-bus contaminants using the advanced methods of data mining and artificial intelligence. The research objective was accomplished by executing the following steps. Firstly, an experimental field program was implemented to develop a comprehensive one-year database of the hourly averaged in-bus air contaminants (carbon dioxide (CO2), carbon monoxide (CO), nitric oxide (NO), nitrogen dioxide (NO2), sulfur dioxide (SO2), 0.3-0.4 micrometer (¿¿¿¿m) sized particle numbers, 0.4-0.5 ¿¿¿¿m sized particle numbers, particulate matter (PM) concentrations less than 1.0 ¿¿¿¿m (PM1.0), PM concentrations less than 2.5 ¿¿¿¿m (PM2.5), and PM concentrations less than 10.0 ¿¿¿¿m (PM10.0)) and the independent variables (meteorological variables, time-related variables, indoor sources, on-road variables, ventilation settings, and ambient concentrations) that can affect indoor air quality (IAQ). Secondly, a novel approach to characterize in-bus air quality was developed with data mining techniques that incorporated the use of regression trees and the analysis of variance. Thirdly, a new approach to modeling in-bus air quality was established with the development of hybrid genetic algorithm based neural networks (or evolutionary neural networks) with input variables optimized from using the data mining techniques, referred to as the GART approach. Next, the prediction results from the GART approach were evaluated using a comprehensive set of newly developed IAQ operational performance measures. Finally, the occupant exposure to in-bus contaminants was determined by computing the time weighted average (TWA) and comparing them with the recommended IAQ guidelines. In-bus PM concentrations and sub-micron particle numbers were predominantly influenced by the month/season of the year. In-bus SO2 concentrations were mainly affected by indoor relative humidity (RH) and the month of (open full item for complete abstract)

Committee: Dr. Ashok Kumar PhD (Committee Chair); Dr. Devinder Kaur PhD (Committee Member); Dr. Cyndee Gruden PhD (Committee Member); Dr. Defne Apul PhD (Committee Member); Dr. Farhang Akbar PhD (Committee Member) Subjects: Civil Engineering; Environmental Engineering; Environmental Health

Doctor of Philosophy, The Ohio State University, 2007, Biomedical Engineering

A multi-paradigm approach integrating three novel computational paradigms: wavelet transforms, chaos theory, and artificial neural networks is developed for EEG-based epilepsy diagnosis and seizure detection. This research challenges the assumption that the EEG represents the dynamics of the entire brain as a unified system. It is postulated that the sub-bands yield more accurate information about constituent neuronal activities underlying the EEG. Consequently, certain changes in EEGs not evident in the original full-spectrum EEG may be amplified when each sub-band is analyzed separately. A novel wavelet-chaos methodology is presented for analysis of EEGs and delta, theta, alpha, beta, and gamma sub-bands of EEGs for detection of seizure and epilepsy. The methodology is applied to three different groups of EEGs: healthy subjects, epileptic subjects during a seizure-free interval (interictal), and epileptic subjects during a seizure (ictal). Two potential markers of abnormality quantifying the non-linear chaotic EEG dynamics are discovered: the correlation dimension and largest Lyapunov exponent. A novel wavelet-chaos-neural network methodology is developed for EEG classification. Along with the aforementioned two parameters, the standard deviation (quantifying the signal variance) is employed for EEG representation. It was discovered that a particular mixed-band feature space consisting of nine parameters and LMBPNN result in the highest classification accuracy (96.7%). To increase the robustness of classification, a novel principal component analysis-enhanced cosine radial basis function neural network classifier is developed. The rearrangement of the input space along the principal components of the data improves the classification accuracy of the cosine radial basis function neural network employed in the second stage significantly. The new classifier is as accurate as LMBPNN and is twice as robust. Next, biologically realistic artificial neural networks are dev (open full item for complete abstract)

Committee: Hojjat Adeli (Advisor) Subjects:

Master of Science in Computer Science, Miami University, 2025, Computer Science and Software Engineering

Computer-aided design (CAD) is widely used for 3D modeling in many technical fields, yet the creation of 2D sketches remains a manual step in typical CAD modeling workflows. Automatically generating 2D sketches can help users in CAD modeling by reducing their workload and by streamlining the design process. While sketches inherently possess a graph structure, with geometric primitives as nodes and constraints as edges, the application of graph neural networks (GNNs) to this domain remains relatively unexplored. To address this gap, we introduce SketchGNN, a graph diffusion model designed to generate CAD sketches using a joint continuous-discrete diffusion process. Our approach includes a novel discrete diffusion technique, wherein Gaussian-perturbed logits are projected onto the probability simplex via a softmax transformation. This enables our model to express uncertainty in the discrete diffusion process unlike traditional methods. We demonstrate that SketchGNN achieves state-of-the-art performance, reducing the Frechet Inception Distance (FID) from 16.04 to 7.80 and the negative log-likelihood (NLL) from 84.8 to 81.33.

Committee: John Femiani (Advisor); Khodakhast Bibak (Committee Member); Karen Davis (Committee Member) Subjects: Artificial Intelligence; Computer Science; Information Science

Master of Science in Computer Science, Miami University, 2024, Computer Science and Software Engineering

Adversarial training (AT) is a defense technique used to increase the robustness of neural networks. AT generates adversarial examples that maximize the loss to the model and then adjusts model parameters to minimize that loss. Previous AT methods typically use only a single attack to perturb adversarial examples that maximize loss, and ignore the roles that different image-classes play in determining final robustness. These techniques are thus unable to properly explore the perturbation space and cannot target specific weaknesses of the training data. As a result, they train models with diminished robustness. This thesis proposes class-based adversarial training, which increases the robustness of AT by using a variety of attacks that target the weakest image-classes of the dataset. We designed and implemented two novel algorithms within this category: the Various Attacks (VA) technique and the Advanced Adversarial Distributional Training (ADT++) technique. Using a novel testing framework created to better examine model robustness across a variety of metrics, we conducted a series of experiments on two benchmark datasets. The results demonstrate the superiority of the VA and ADT++ frameworks over state-of-the-art adversarial training methods.

Committee: Samer Khamaiseh (Advisor); Honglu Jiang (Committee Member); Hakam Alomari (Committee Member) Subjects: Computer Science

Doctor of Philosophy, The Ohio State University, 2024, Animal Sciences

The urgent shift from fossil fuels to renewable energy sources highlights the critical need for innovative and sustainable biofuel production technologies. However, a significant hurdle in biofuel production especially butanol, pentanol, hexanol, heptanol, and octanol is the toxicity of these compounds to microorganisms. Extensive process engineering efforts, including vacuum-assisted gas stripping (VAGS), have been made towards in-situ recovery of butanol to alleviate the problem of product toxicity to producing microorganisms. Despite the success of VAGS in butanol recovery, it is still riddled with the problem of excessive water removal from the bioreactor during product recovery. The ongoing use of superhydrophobic separation materials for oil recovery in oil spillage situations indicated that there could be a way to improve water/butanol separation during in-situ recovery. Thus, this study explored the use of superhydrophobic separation materials within a VAGS system for engendering significant water retention within a bioreactor and enhancing efficient recovery of biofuels, specifically butanol and high molecular weight alcohols (C4 – C8), from fermentation broths. Central to this investigation is the development of superhydrophobic and superoleophilic stainless steel meshes (SSM) using polydimethylsiloxane (PDMS) and polytetrafluoroethylene (PTFE). The SSM, having a water contact angle of 156.48° was incorporated into the VAGS setup and used to enhance the recovery efficiency and economic feasibility of the biofuel production process. Experimental and modeling approaches including the use of artificial neural networks (ANN) modeling were employed to optimize the recovery conditions and assess the interplay between process parameters and system performance. Thus, the study explored the influence of several critical parameters, including mesh pore size, vacuum time, initial alcohol concentration, and bioreactor operational conditions, on the performance of th (open full item for complete abstract)

Committee: Thaddeus Ezeji (Advisor); Victor Ujor (Committee Member); Ajay Shah (Committee Member); Gonul Kaletunc (Committee Member); Alejandro Relling (Committee Member) Subjects: Animal Sciences; Energy; Engineering

Master of Computing and Information Systems, Youngstown State University, 2024, Department of Computer Science and Information Systems

The field of coding education is rapidly evolving, with emerging technologies playing a pivotal role in transforming traditional learning methodologies. This thesis introduces Dr. Lego, an innovative framework designed to revolutionize the assessment and understanding of block-based coding through the integration of sophisticated deep learning models. Dr. Lego combines cutting-edge technologies such as MobileNetV3 (Howard, 2019), for visual recognition and BERT (Devlin et al., 2018), and XLNet (Yang et al., 2019) for natural language processing to offer a comprehensive approach to evaluating coding proficiency. The research methodology involves the meticulous curation of a diverse dataset comprising projects from the LEGO SPIKE app (LEGO Education, 2022), ensuring that the models are subjected to a broad range of coding scenarios. Leveraging the dynamic educational environment provided by the LEGO SPIKE app (LEGO Education, 2022), Dr. Lego empowers users to design and implement various coding projects, fostering hands-on learning experiences. This thesis delves into methodologies aimed at enhancing coding education by exploring model integration, data generation, and fine-tuning of pre-trained models. Dr. Lego not only evaluates coding proficiency but also provides cohesive and insightful feedback, enhancing the learning experience for users. The adaptability of the framework highlights its potential to shape the future of coding education, paving the way for a new era of interactive and engaging learning experiences.

Committee: Abdu Arslanyilmaz PhD (Advisor); Feng Yu PhD (Committee Member); Carrie Jackson EdD, BCBA (Committee Member) Subjects: Computer Science; Engineering; Information Systems; Robotics; Teaching

PhD, University of Cincinnati, 2023, Engineering and Applied Science: Computer Science and Engineering

Fingerprinting is a technique that identifies websites, software and devices by leveraging a group of information from users. An attacker can acquire users' secrets by only analyzing side-channel features from a system, such as network packet size and direction, power usage or CPU usage. In traditional fingerprinting attacks, a large amount of human effort is required as an attacker has to manually extract effective features for attacking purpose. This kind of attacking approach is easy to be defended as a defender can invalidate the attack by modifying the target features that are used in the attack. However, for AI-based fingerprinting, handcrafted feature is not necessary anymore. An attacker can train a machine learning classifier over raw data directly and achieve an impressive classification results. In this proposal, I propose to design effective and efficient defenses against deep neural network based fingerprinting attack. Firstly, I propose to improve the efficiency of existing defense against neural network based stream fingerprinting. Many defense algorithms have been proposed to defeat stream fingerprinting. However, most of those existing defense algorithms need extremely high bandwidth overhead in order to make the defense effective. In this dissertation, I leverage feature selection methods to analyze the feature space in stream fingerprinting. Instead of treating network packets equally when adding noises, we distinguish important packets using feature selection algorithms and add more noise to those important packets. Secondly, I propose to design an efficient defense against CPU cache based website fingerprinting. Recently, a new attack was proposed to monitor the cache occupancy of the Last Level Cache on a user's CPU. Although a defense was proposed, it is not effective when an attacker adapts the classifier with defended data. In this dissertation, I investigate the behavior of cache occupancy channel and reveal the reaso (open full item for complete abstract)

Committee: Boyang Wang Ph.D. (Committee Chair); Nirnimesh Ghose Ph.D. (Committee Member); Tingting Yu Ph.D. (Committee Member); Nan Niu Ph.D. (Committee Member); Wen-Ben Jone Ph.D. (Committee Member) Subjects: Computer Engineering

Master of Science, University of Toledo, 2023, Civil Engineering

Beam-column joints play a critical role in transferring forces between beam and column elements and maintaining structural integrity during severe loading. While the nonlinear behaviors of beams and columns are commonly modelled in global frame analyses through the use of plastic hinges, the behavior of joints through the use of rigid end offsets is often omitted. The objective of this study is to develop an artificial neural network and derive the plastic hinge curves required for modeling beam-column joints in global frame analyses. As the first step, a feed-forward artificial neural network (FFNN) is developed to predict the shear strengths of beam-column joints. A comprehensive dataset of 598 experimental joint specimens is compiled from 153 previously published research studies. The 555 data points which passed the exploratory data analysis are used to train, test, and validate the proposed network for applicability to a wide range of input variables and joint configurations. The accuracy and reliability of the proposed FFNN were evaluated using a comprehensive set of evaluation metrics in comparison with three existing networks from the literature. The network predicted shear strength is used to derive shear stress-strain and moment-rotation curves for joint hinges. A spreadsheet tool is developed to execute the network formulations, calculate joint shear strength, and derive joint hinge curves for practical use by engineers and researchers.

Committee: Serhan Guner (Committee Chair); Luis Alexander Mata (Committee Member); Douglas Karl Nims (Committee Member) Subjects: Civil Engineering

Master of Science, Miami University, 2023, Computer Science and Software Engineering

Traditionally, software modeling has relied on conventional input devices such as keyboards and mice. However, as new interaction methods become more popular, development environments must adapt to these evolving needs. Moreover, nontraditional interfaces offer the potential for improved accessibility. This thesis introduces an innovative framework for intelligent voice-driven software modeling. The framework leverages advanced technologies, including speech-to-text conversion, natural language processing, and domain-specific input commands. By combining these elements, this research presents a powerful and intuitive system that allows users to create software models through voice commands. The framework's effectiveness has been evaluated primarily via two different user studies and secondarily using cross-validation to evaluate trained machine learning models. The evaluation of the final implementation of the framework resulted in an average command accuracy of 79.4% and an average overall rating of 7.85 out of 10 from the participants. Overall, this study demonstrates the viability and potential of voice commands as an effective interface for software modeling. By embracing voice-driven interactions, this thesis aims to improve accessibility, user experience, and overall efficiency in software engineering.

Committee: Eric Rapos (Advisor); Christopher Vendome (Committee Member); Xianglong Feng (Committee Member) Subjects: Computer Science; Engineering

Doctor of Philosophy, The Ohio State University, 2023, Electrical and Computer Engineering

The advent of deep neural networks (DNNs) have brought exciting new possibilities for the realization of automated driving functions. These data-driven methods have been widely applied to various driving tasks, including end-to-end urban driving. However, the use of these methods beyond simulated tests remains limited due to two significant shortcomings: (i) the lack of model transparency and (ii) the difficulty of generalizing beyond the training distribution. This dissertation aims to investigate methods for addressing the transparency and safety mitigation of learning-based controllers, specifically deep reinforcement learning (DRL) methods, to enable safe and predictable driving. To enhance interpretability, an interpretable and causal state representation, coined the driving forces, is proposed. This representation captures the causal relationship between the state and the produced control action by leveraging force features to encode the influence of internal and external factors on the ego vehicle. By training a DRL agent on this representation within a highway driving environment, the ability of the driving forces to encode and interpret the state-action causalities was demonstrated. Furthermore, an alternative paradigm for online adaptation by modifying the formulation of the driving forces is proposed to mitigate the behavior of the ego vehicle. The results showed that the ego vehicle was successful in mitigating its behavior and following desired new and unseen behaviors, without requiring modification to the underlying DNN. To address model transparency in black-box DNN-based driving policies, a knowledge distillation framework that combines interpretable decision trees with rule learning algorithms is proposed. This framework learns decision rule sets that represent the decision boundaries of the original driving policy. The driving forces are utilized to abstract the original state representation and ensure the interpretability of the learned explanati (open full item for complete abstract)

Committee: Umit Ozguner (Advisor); Keith Redmill (Committee Member); Qadeer Ahmed (Committee Member); Gladys Mitchell (Committee Member) Subjects: Artificial Intelligence; Computer Engineering; Electrical Engineering; Transportation

Master of Science in Engineering, University of Akron, 0, Mechanical Engineering

The online monitoring of the lubricant oil is critical to the health status of machinery in automotive, power generation, agricultural, and various other industries. Proper lubrication allows machinery to operate efficiently, effectively, and is essential to preventing machine failure and costly repairs. There are multiple critical properties of a lubricant oil, such as total acid number (TAN), total base number (TBN), water content, soot concentration, diesel contaminant concentration, etc, which can be used to characterize the health status and effectiveness of a lubricant oil. Various sensors exist to monitor these lubricant properties individually, but are either limited by slow responses, need for expert analysis, inadequate accuracy, or suffer from cross sensitivity issues. Even though these methods exist, there is still a need for a sensing system capable of accurately monitoring multiple properties in real time. Artificial neural networks (ANNs) have been used with sensing arrays to overcome cross sensitivity issues and create a rapid method of data analysis from the sensor response. These ANNs are extremely beneficial towards the lubricant oil monitoring process but have shown limitations due to several issues. Firstly, typical ANNs will require a large amount of samples for the training process. In the case of lubricant oil monitoring, this means the tedious and lengthy testing and creation of many samples with various levels of property concentrations. Secondly, the establishment of an ANN requires expertise and research to decide the type of ANN and to avoid errors in the training process. Underfitting and overfitting are two common issues that can arise from network training and can both result in large errors when testing the system. To address these above issues, in my thesis two sensor arrays were developed to monitor multiple lubricant oil properties. First, a microsensor array was developed to monitor (open full item for complete abstract)

Committee: Jiang Zhe (Advisor); Amir Nourhani (Committee Member); Ge Zhang (Committee Member) Subjects: Mechanical Engineering

BA, Oberlin College, 2022, Computer Science

In open multiagent systems, multiple agents work together or compete to reach the goal while members of the group change over time. For example, intelligent robots that are collaborating to put out wildfires may run out of suppressants and have to leave the place to recharge; the rest of the robots may need to change their behaviors accordingly to better control the fires. Thus, openness requires agents not only to predict the behaviors of others, but also the presence of other agents. We present a deep reinforcement learning method that adapts the proximal policy optimization algorithm to learn the optimal actions of an agent in open multiagent environments. We demonstrate how openness can be incorporated into state-of-the-art reinforcement learning algorithms. Simulations of wildfire suppression problems show that our approach enables the agents to learn the legal actions.

Committee: Adam Eck (Advisor) Subjects: Artificial Intelligence; Computer Science