2021 Capstone Showcase: May 6 Virtual Capstone Event

Team: Chuzhen Wang [IEOR], Hanyu Xia [IEOR], Alexis Ruiz [ME], Alex Ko [IEOR], Shan Lu [IEOR]
Advisor(s): Damien Thioulouse [Asurion/Simplr], Clément Ruin [Asurion/Simplr], Harsh Tomar [Asurion/Simplr], Lee Fleming [IEOR]

The large amount of customer service data presents enormous opportunities, particularly in the Natural Language Understanding (NLU) domain. Our team explores a wide range of customer service data from retail, software to food industries. We apply different NLP techniques and machine learning models to build an AI-based parser that is able to understand email messages efficiently and connect customer’s intents with product items from customer tickets.

Team: Charlotte Lao [BIOE], Saahil Patel [BIOE]
Advisor(s): Todd Roberts [ATDEV], Grace O’Connell [ME]

Muscular dystrophy causes degradation of muscle mass over time, and life-long physical therapy is the only solution to rehabilitate and strengthen muscles. Unfortunately, maintaining a constant in-person physical therapy regimen is difficult due to the lack of accessibility, lifetime cost, and limited mobility of these patients. An at-home physical therapy device by Assistive Technology Inc. exists, and we are improving on the device’s alignment to the arm, since misalignment can cause incorrect exercise form and discomfort for the user. Our implementation includes Inertial Measurement Units to monitor the joint angle of the elbow with the device’s built-in encoder.

Team: Ada Sun [IEOR], Cara Zhu [IEOR], Seth Meng [IEOR]
Advisor(s): Lei Chen [Bloomberg], Dave Lingenbrink [Bloomberg], Tarek Zohdi [ME]

Bloomberg obtains trading data from banks around the world. Our team started with “over-the-counter” asynchronous time series data. We sought to represent separate signals as a one signal with all dimensions and duration indicators as additional features. We then worked at implementing Significance-Offset Convolutional Neural Network (SOCNN) architecture with asynchronous time series using real-world scenarios. Our model takes input from OTC data (based on the derivatives market) and forecasts the future prices. We also evaluate the model performance with other deep learning models such as CNN, Phase LSTM, and RNN. Our proposed model can provide Bloomberg clients with better derivatives investment strategies and a comprehensive evaluation of the potential market.

Team: Derek Ho [ME], Calvin Shih [ME], Anna Wolfe [ME], Xingshuo Yan [CEE]
Advisor(s): Daniel Daugherty [Blue Goji], Austin Peck [Blue Goji], Siyuan Ren [Blue Goji], Coleman Fung [Blue Goji], Gabriel Gomes [ME]

An estimated 13 million people are currently living with an adult-onset brain disorder, such as Alzheimer’s disease, stroke, and Parkinson’s disease in the US. These individuals are more prone to falling, and the rehabilitation of declining balance and postural stability is a challenging process. With the Blue Goji Infinity system, we have created a dynamic testing environment that utilizes VR gaming to direct users with tasks that challenge balance and posture. Using gameplay data synchronized with movement data collected through force sensors and a depth camera, we are able to track and improve the rehabilitation of neurological disorders.

Team: Zhecan Huang [IEOR], Camilla Nawaz [IEOR], Zehao Wei [IEOR], Xuanrui Zhang [IEOR]
Advisor(s): Anne Spitz [Crossing Minds], Paul Grigas [IEOR]

Recommendation systems are machine learning models for suggesting relevant items to users, and they are the core product of tech companies, such as Netflix, Amazon, and Crossing Minds, etc. However, training and choosing the best recommendation models are resource-intensive and time-consuming processes because of the wide range of pre-setting options for models. Our team is currently creating an Automated Machine Learning pipeline to generate reduced searching ranges for pre-setting options, known as hyperparameters. We aim to use the reduced ranges with Bayesian optimization to tune the hyperparameters of recommenders and thus save computational time and money.

Team: Andrew Lee [MSE], Malo Le Magueresse [IEOR], Shiyan Xu [ME], Xinyue Xu [ME], Yun-Yi (Wayne) Chu [MSE], Ziren Li [IEOR]
Advisor(s): Leila Teichmann [Gerdau], Rodrigo Proença [Gerdau], Belisa Bunonafina [Gerdau], Fernanda Bordin [Gerdau], Tarek Zohdi [ME]

The objective of this project is to create a machine learning model to predict scrap and rework in steel manufacturing. This will help our client anticipate production delays and reduce financial losses.

Team: Shae Alhusayni [IEOR], Chris Lai [IEOR], Hong Tang [BIOE], Charles Cui [IEOR]
Advisor(s): Sarine Babikian [Glooko], Paul Grigas [IEOR]

Diabetes, the current 7th leading cause of death, is projected to increase by over 50% by the end of this decade. Since it is an incurable disease, the only way to alleviate symptoms is by managing the disease closely. Our team’s objective is to implement machine learning algorithms that will help patients using Glooko’s diabetes management app obtain insights on how to manage their diabetes more effectively. Specifically, our machine learning solution will be split into two parts: 1- apply regression and classification models to predict users’ diabetes outcomes; 2- apply natural language processing techniques on user-entered text to understand and improve user experience

Team: Katie Henshaw [ME], Matthew Mesman [ME], Julie Yu [ME]
Advisor(s): Jack Miller [Move2Play], Brenden McMorrow [Move2Play], Tarek Zodhi [ME]

Kids spend on average 7.5 hours a day watching a screen. This unprecedented amount of screentime can lead to a range of health issues including increased obesity, decreased academic performance and decreased mental health outcomes. To combat this, Move2Play is designing toys to encourage kids to be active. Our team is working to add to their portfolio with a new bubble toy. Specifically, we are utilizing fluid dynamics and engineering principles to optimize the components of the toy to create bubbles far larger and longer lasting than what is currently available on the market.

Team: Abhishrut Sinha [IEOR], Qiuyu Liu [IEOR], Tingting Chen [IEOR], Yushan Liu [IEOR]
Advisor(s): Pablo Mastretta [MSCI], Cloud Force [MSCI], Lee Fleming [IEOR]

To expedite the research to market process, financial researchers need more efficient and scalable ways to analyze financial datasets, construct factor models and get client feedback. At the same time, there is a need for modernized client engagement tools that allow interactivity and visual clarity.

The new model dashboard will be a part of the new factor model engine for predictive analysis and risk control that brings research to market rapidly in an easy-to-configure workflow system. We’re utilizing python’s analytical libraries (xarray, pandas, ipywidgets) & MSCI datasets for both Equity and fixed income models to build this platform. We aim to help MSCI Production Services, Product Owners, Researchers and MSCI stakeholders by modernizing our engagement tools using a state of the art dashboard to clearly show the power of our models. The dashboard also improves the understanding of the finance models for product managers and business analysts, so they could answer client questions easily, accelerate the feedback cycle and improve the research to market process.

Team: Sarah Gunasekera [ME], Bogdan Cristei [IEOR], Elliott Suen [CEE], Sydney Holgado [CEE]
Advisor(s): Seth Hoedl [Post Road], Gabriel Gomes [ME]

Global energy demand is expected to rise by 50% by the year 2040 – the energy grid of the future calls for decarbonization, digitalization and decentralization. We must transition away from a centralized energy system that relies on fossil fuels to a more distributed system that prioritizes renewables. Our team built a tool that uses a data driven approach to find optimal modernized grid configurations in order to ensure rapid and cost-effective decarbonization.

Team: Joey Zhu [ME]
Advisor(s): Alice Agogino [ME], Mark Mueller [ME], Clark Zha [ME], Doug Hutchings [Squishy Robotics]

Quadcopters and other flying robots are typically fragile and unable to operate in rough, unknown environments typical of disaster zones. Our team has designed and assembled a collision resilient flying robot that utilizes a rigid tensegrity shell and can carry a wide range of payloads, including the heritage Squishy Robotics HazMat payload. The rigid icosahedron tensegrity shell, mounting apparatus, and control system allow the quadcopter to continue operations even after a 6.5 [m/s] collision.

Team: Celina Chow [ME], Toby Liu [ME], Tamara Perrault [ME], Edward Xia [ME]
Advisor(s): Alice Agogino [ME], Doug Hutchings [Squishy Robotics]

First responders do not currently have access to technology which allows them to prepare a response to hazmat scenarios while staying a safe distance away. Squishy Robotics is creating a robot that can be accurately dropped from a drone into a disaster area and send relevant data to responders. A 3D printed flexible payload and outer tensegrity structure paired with precise wind-measuring technology allow the robot to effectively perform remote chemical sensing. The result is an accurately dropped, waterproof, sensing robot that can survive impact on any terrain and help first responders generate a more informed, and thus safer response.

Team: Benjamin Boggs [ME], Daniel Newell [ME], Mina Fanaian [ME]
Advisor(s): Alice Agogino [ME], Doug Hutchings [Squishy Robotics]

Every year there are 500,000 calls to US fire departments for emergencies involving hazardous materials, where first responders must don hazmat suits for hours to contain the situation. Our project aims to develop mobile capabilities for an air-droppable sensor payload to reduce firefighters’ hazmat exposure. The robot is robust to impacts due to its tensegrity construction. By adjusting cable lengths strategically, we can effectively shift the robot’s center of mass, enabling a series of punctuated movements that allow it to map the extent of the chemical spill. This data can be sent offsite, reducing the need for on-site personnel.

Team: Chen Guo [BIOE] Arsany Gad [BIOE] Dechao Zeng [ME]
Advisor(s): Dave Fabry [Starkey], Tarek Zohdi [ME]

Tinnitus is a disorder that affects up to 25% of Americans. It is characterized by experiencing ringing or other kinds of noise in the ear, and it adversely affects stress levels and mental health. Sound therapy is the most promising treatment to reduce symptoms of tinnitus. The Happy Hearing team is designing a stress classifier for feedback control of tinnitus treatment sound filters that work in conjunction with established presets. Our team’s solution will include a machine learning algorithm that can process biosignals, categorize stress levels, and adjust the tinnitus sound filter that is a part of Starkey’s hearing aid, the Livio Edge AI.

Team: Tuan Tu [ME], Jing-Song Huang [ME], Mohammed Matar [ME], Mohammad Iravanian [ME]
Advisor(s): Homayoon Kazerooni [ME]

Stairs are widely prevalent in human environments, but require considerable energy, strength, and coordination effort to traverse. One way to solve this is by using exoskeletons. Current exoskeletons can provide assistance during level walking only, so we are implementing an add-on to current exoskeletons to enable stair detection and traversal. We use walking cycle analysis and machine learning algorithms to detect the user’s action and provide torque to traverse stairs with less effort, giving them more accessibility.

Team: Nicholas Callegari [ME], Yi Zeng [ME], Zejun Zhong [ME], Zhouyang Qi [ME]
Advisor(s): Homayoon Kazerooni [ME]

Around 69% of the US labor force experiences workplace fatigue, with about 30% of all workplace injuries resulting in lost workdays due to repetitive stress injuries. Our team is working on a new iteration of an actuator that will be located on the hip of various exoskeleton devices, reducing the strain on users when executing repetitive movements involving the legs and back. We are implementing a torque sensor into the actuator to effectively help detect user inputs in order to achieve zero-impedance for routine movements like walking. This new actuator will significantly reduce the prevalence of workplace repetitive injuries.

Team: Bo Zhou [BIOE], Sixtine Lauron [IEOR], James Corbitt [BIOE], Samuel Harreschou [NE]
Advisor(s): Gabriel Gomes [ME], Gundolf Schenk [UCSF]

Electronic health records consist of text-heavy clinical notes that document disease history of patients : this unstructured data makes it very challenging to extract information and examine medical trends. Our solution? By applying NLP and deep learning on hospital notes, we created a patient profiling search engine : our web-based user interface allows physicians to study for each new patient the medical history and medication reactions of its N nearest neighbors in terms of symptoms, lab results, diagnoses and medications. By taking a glance at our 3D patient clusters, doctors can adjust their medical prognosis and preferable treatments decisions.

Team: Himani Patel [ME], Sami Lama [BioE]
Advisor(s): Grace O’Connell [ME], Robert Matthew [UCSF]

There are 1.6 million amputees in the US and current socket-type prostheses have inadequate clinical outcomes. Osseointegration, a promising new technology, fuses the implant directly into bone to improve quality of life. However, the emergence of cracks in surrounding bone requires further investigation of biomechanical loading. Our team is developing a model that analyzes and predicts risk of fracture in patients with osseointegrated prosthetics. With tools like Matlab and Solidworks, collected patient data is used to calculate forces acting on the implant. Activities that pose higher risks can then be identified to provide guidelines for patients with osseointegrated prostheses.

Team: Luke Murdough [ME], Christopher Scaduto [ME], Hanyi Lu [MSE], Stone Chen [MSE], Xiaoyun Li [MSE]
Advisor(s): Matthew Sherburne [MSE], Peter Hosemann [NE]

Figuring out the right metal 3D printer for your needs is a time-consuming and confusing process as you look through scores of webpages of companies to start figuring out the one to suit your needs. The goal of our project is to design and implement a more user-friendly web app for selecting a 3D printing method to meet the demand of the customer/researcher. By organizing existing research on metal 3D printers and narrowing the selection criteria to only the most important component features, we hope to create an easy to use website that will satisfy any researcher.

Team: Liu Rui [EECS], Songwen Su [EECS], Yifei Zhang [EECS]
Advisor(s): Dawn Song [EECS], Peng Gao [EECS]

There is a cyber attack happening every 39 seconds. But many people and companies do not have enough knowledge about the current trends or behaviors of cyber attacks and how to prevent them. Our team is developing a security domain question answering system that aims to promote cybersecurity knowledge sharing. Our solution is a web page user interface that uses both Bidirectional Encoder Representations from Transformers (Bert) language model and knowledge-graph based approach to provide users with easy access to precise answers of security questions.

Team: Lisa Qing [EECS], Xuewei Wang [EECS], Xianlin (Shay) Shao [EECS]
Advisor(s): Brian A. Barsky [EECS]

In 2019, the World Health Organization recorded at least 2.2 billion people with a vision impairment, such as nearsightedness and farsightedness. VCDs use both hardware and software to generate clear displays corresponding to the user’s eye prescription through image reverse-blurring algorithms. However, current VCD algorithms only allow single-eye viewing from a fixed position. Our team aims to eliminate these limitations by enabling screen devices’ built-in cameras for eye-tracking to allow dynamic viewing positions and introducing parallax barrier technology to incorporate two-eye viewing. This project will validate the practicality of VCDs by bringing impractical state-of-art algorithms to real-world usage settings.

Team: Junhao Yu [ME], Zhanqin Huang [ME]
Advisor(s): Tarek Zohdi [ME]

Real-time mapping of complex environments after multi-location wildfires cannot be completed by traditional remote sensing methods like satellites and aircrafts. These methods are either slow, tricky to deploy, or have limited flexibility. Our drone-swarm mapping solution combines Machine Learning and commercial drones to implement a competitive alternative. The result is an efficient, adaptive, and cost effective solution that can track and predict wildfire progression.

Team: Gwenlyn Angel Cabiltes [MSE], Jonathan Chow [MSE]
Advisor(s): Aaron Hall [MSE], Ting Xu [MSE]

The world faces a severe plastic waste problem: Nearly 80% of plastic products end up in landfills or the environment. Biodegradable plastics like polylactic acid (PLA) are a promising solution but face extensive difficulties degrading and composting in most facilities. Our group addresses this issue by pinpointing economically and technologically viable solutions which harness the potential of enzymatic degradation to convert back into simple monomers, producing minimal waste while maintaining material properties. Embedding the degrading enzymes in products enables quicker and thorough breakdown of plastic with the aid of water and heat, and is also compatible with existing composting and waste infrastructure.

Team: Zachary McGuire [NE], Daniel Payne [NE], Andria Sperry [NE], Justin Gonzalez [NE], Nebeyu Yonas [NE]
Advisor(s): Ed Morse [NE]

The interaction between plasma and water can eliminate even the toughest of hazardous organic and inorganic compounds in water left over from municipal treatment. Our project endeavors to deliver a concept design and theoretical throughput analysis of an efficient, affordable device for household use. We hope to increase consumer access to this technology and inspire a continuation of this work.

Team: Armand Sauzay [IEOR], Yiman Hu [IEOR], Zhihan Hui [IEOR], Ziwen Jiang [IEOR], Shengfeng Li [IEOR], Vijaykumar Vishnu Swaroop [IEOR]
Advisor(s): Paul Grigas [IEOR]

Online advertising is a multi-billion dollar industry and is only poised to grow with time. Advertisers bid for ad slots based on how well they can profile users on the internet. This comes at a cost to user privacy. Using open source datasets and machine learning packages, we will first see how advertising agencies bid the ads, and then see how advertisers can bid for slots as effectively without using sensitive attributes which compromise user privacy. Our project can provide insights to a new bidding/design mechanism that agencies can use to achieve their goal while preserving users’ privacy.

Team: Lucy He [IEOR], Liwen Zhang [IEOR], Mathilde Bachy [IEOR], Sean Wang [IEOR] and Klara Guan [IEOR]
Advisor(s): Zeyu Zheng [IEOR]

Nowadays, logistics operations are run entirely digitally, allowing organizations to make significant business decisions based on data. With logistic operations analysis, online shopping businesses can facilitate quick and live feedback loops to monitor their networks. The reality is that customer satisfaction has become a hot topic and a competitive advantage for organizations. Online Shopping platforms thrive on network effects and could not survive without their user community’s approval and referrals. Therefore, we work to identify actionable insights in order to improve customer reviews based on the prediction results of machine learning models.

Team: Frank Cai [EECS], Sihao Chen [EECS], Sophie Wu [EECS], Weili Liu [EECS], Xuantong Liu [EECS], Yizhou Wang [EECS]
Advisor(s): Brian Barsky [EECS]

Individuals with severe motion impairments may not be able to control a mouse or type on a keyboard. Unfortunately, there are very few viable alternatives to a computer mouse right now. To address this problem, our team is working on developing a computer-vision based input system with a webcam as its input device. This system will be operating on low-cost devices to help people with difficulty using computer mice. We propose different methods for mapping arbitrary hand configurations to discrete gestures. This allows a user to fully emulate the actions of a computer mouse without any physical interaction.

Team: Alan Liang [EECS]
Advisor(s): Alvin Cheung [EECS]

Automatically grading student answers is often a necessity for large courses that have assignments in SQL, a programming language used by most database systems. Existing solutions have focused on grading the output of a SQL query rather than the query itself. Team Alan leverages Cosette, an automated SQL query equivalence prover, to implement autograding. We build a SQL parser and add additional functionality on top of Cosette to evaluate correctness. By integrating Cosette into Otter, an open source autograder, we implement our solution into the SQL lab for a data science course at UC Berkeley with more than 1000 students.

Team: Shubha Jagannatha [EECS], Maggie Zhang [EECS], Weiyan Zhu [EECS]
Advisor(s): Allen Yang [EECS]

Beyond-line-of-sight drone navigation is currently unsafe due to the absence of tools to accurately perceive and understand a drone’s surroundings. Our team is working on visualizing a drone’s environment through Virtual Reality in real-time, making remote drone piloting much easier and safer. The solution we present is an end-to-end pipeline for remote piloting using a drone-mounted stereo camera, computer vision algorithms, and a Unity visualizer.

Team: Ashank Verma [EECS], Bo Pang [EECS], Jingchao Zhou [EECS]
Advisor(s): Inigo Incer [EECS], Alberto Sangiovanni Vincentelli [EECS]

Cyber-Physical Systems (CPS) are important components of critical infrastructure and must operate with high levels of reliability and security. However, conventional CPS are not built with adequate security measures to withstand modern cyberattacks. Our team developed a Cyber-Physical Immune System (CPIS) that can be deployed onto a conventional CPS and provides extra protection. Inspired by its biological cousin, CPIS consists of distributed computing units with independent networks, utilizes unsupervised machine-learning algorithms to identify anomalies in the conventional CPS, all while adapting to the changing environment. If threats are detected, CPIS raises awareness and puts the victim into safe mode to minimize the effects of cyberattack.

Team: Albert Loekman [EECS], Christina Lu [EECS], Haochong Xia [EECS]
Advisor(s): Alex Bayen [EECS], Jessica Lazarus [EECS]

According to the San Francisco County Transportation Authority, car trips are expected to grow by 36% from 2015 to 2050, disproportionally increasing congestion and travel costs for low-income communities. Through BISTRO, we provide an open-source traffic simulation tool and graphical dashboard for policymakers to make traffic policies on how to combat these issues. To do this, we calibrated a traffic simulator for San Francisco and developed an algorithm to find the best locations and prices for a toll zone based policy. The dashboard will show how traffic conditions improve after implementing these toll zones.

Team: Zoey Liaowang Zou [EECS], Zecheng Wu [EECS]
Advisor(s): Alberto Sangiovanni Vincentelli [EECS]

Currently, 93 million people worldwide live with diabetic retinopathy (DR). DR is a well-known sight-threatening complication of diabetes mellitus,which can cause blindness if left undiagnosed and untreated. Our team is working on a machine learning algorithm that improves disease screening efficiency, especially for early-stage cases. The resulting model can maximize detection performance by referring most uncertain cases to doctors, cutting screening cost and improving diagnostic accuracy.

Team: Cem Koc [EECS], Eric Liu [EECS], Jiayi Wang [EECS], Yujie Xu [EECS]
Advisor(s): Alvin Cheung [EECS]

The ongoing AI arms race sparked by Krishevsky et al. in 2012 with breakthrough artificial deep neural network performance has taken the world by storm. As AI has advanced, there has also been massive growth in AI frameworks and tools. Unfortunately, learning how to use these deep learning frameworks has been a challenge for engineers, researchers, and laity due to a steep learning curve. With our research, we aim to increase accessibility and productivity of researchers, industry professionals, and tech-curious people in machine learning by providing a data-driven model for searching, generating, and summarizing AI code.

Team: Chet Singh [IEOR]
Advisor(s): Lee Fleming [IEOR]

VideoPlus, a video surveillance product, can neither detect stationary targets nor classify the moving targets it detects. This project aims to upgrade VideoPlus, transforming the current product into a state-of-the-art surveillance system capable of detecting both stationary and moving objects as well as characterizing the detected objects as either objects of interest (e.g., vehicles, humans) or non-interest (e.g., animals, vegetation). By incorporating deep-learning-driven artificial intelligence, the enhanced surveillance system will deliver superior performance in real time, be robust across disparate data domains, and enable cross-product application.

Team: Ricardo Juarez Martinez [BIOE]
Advisor(s): Syed Hossainy [BioE] and Dorian Liepmann [BioE]

Osteoarthritis (OA) affects millions of people and is the leading cause for disability in the US. As the disease progresses and cartilage is damaged, the standard of care is reactive and invasive. Our team is paving the way for a regenerative approach that can roll-back the progression of OA. By combining an injectable electro-active polymer with an external electromagnetic field our team hypothesizes that an electrical response will be achieved. To prove this, our team is creating a reliable test method to characterize the electrical behavior of our system. If responsive, this system can be used to assess cartilage regeneration in-vitro.

Team: Jingwei Liu [IEOR], Zongyang Li [IEOR], Huiyi Qi [IEOR], Luke Chen [IEOR]
Advisor(s): Xin Guo [IEOR], Lizeng Zhang [IEOR]

In high-frequency trading, traders need to place large orders under a specific time frame. When trading, they need to choose between the choice of limit orders, which guarantee the price at execution but have a risk of not being executed on time, or market orders, which guarantee immediate execution but have the risk of paying a higher price. Traders often face the dilemma between bearing higher execution risk or higher trading costs. Our team’s objective is to reduce trading costs under an acceptable execution risk by improving existing reinforcement learning models to figure out how to place orders optimally.

Team: Zhenghang Xie [IEOR], Qi Chen [IEOR], Di Kan [IEOR], Yizhuo Li [IEOR], Zexin Xia [IEOR]
Advisor(s): Lizheng Zhang [IEOR], Svitlana Vyetrenko [IEOR], Xin Guo [IEOR]

Using AI models to make optimal trading decisions is gaining more popularity in quantitative investment. Our team aims to maximize revenue and minimize transaction costs based on reinforcement learning models. To achieve this goal, we built various RL models such as DQN and DDQN. All models are trained by previous real order data and tested by simulation platforms. The results can be promising to support brand-new high frequency trading strategies and investigate deep principles about the stock market’s micro-structure.

Team: Kexin Fang [IEOR], Zhiyang Han [IEOR], Zhangguanglu Wang [IEOR], Xinzhou Xu [IEOR]
Advisor(s): Xin Guo [IEOR], Lizeng Zhang [IEOR]

With an average daily trading volume of 10.9 billion dollars, the stock market has always been a roller coaster where panics and opportunities both exist, especially in the special year of 2020. Our team aims to make accurate predictions on future trends of stocks for investors to make winning trading strategies and reduce risks in times like this. We would use the technology of multi-agent reinforcement learning to simulate the changes of stock prices and present real-time forecasts based on historical market data.

Team: Logan Schultheis [IEOR], Sylvia Gu [IEOR], Xingchen Liao [IEOR], Maggie Yeh [IEOR]
Advisor(s): Xin Guo [IEOR], Lizeng Zhang [IEOR]

High-frequency trading has now accounts for around 70% of the US equity trading. However, the optimal placement problem, one of the key problems in this field that aims to find the optimal trading price in small time scale, has much less practical solutions than others. Our team is using reinforcement learning, which trains intelligent agents to take correct actions to maximize cumulative rewards, to build an automatic trading algorithm to help make the optimal trading decisions. This algorithm will effectively maximize total trading profits and beat most of the experienced traders in the market.

Team: Yuhang Feng [IEOR], Tong Zhou [IEOR], Zeyu Li [IEOR], Zhao Zou [IEOR]
Advisor(s): Lizeng Zhang [IEOR], Xin Guo [IEOR]

In today’s financial market, there is a high demand and practical use of high-volume optimal order execution in investment decisions and high-level trading strategies. Therefore, it is essential to have a realistic simulated market where different execution strategies can be tested repeatedly to minimize the risk. Our project’s goal is to simulate such a realistic market containing agents that reflect on real-time price changes and the simulation could be used to test different trading strategies. The whole simulation environment is based on the ABIDES algorithm, introduced by David Byrd. We will validate our system with Nasdaq ITCH-5.0 historical data.

Team: Ziyue Yu [IEOR], Hongye He [IEOR], Yuying Chen [IEOR], Yichen Yuan [IEOR], Zhengdong Li [IEOR]
Advisor(s): Xin Guo [IEOR], Lizeng Zhang [IEOR]

Corporate bonds’ limited liquidity and intransparent market condition result in an unfriendly bond transaction environment, especially for individual investors. Our team intends to build a benchmark for the corporate bonds’ transaction cost analysis, facilitating investors to evaluate their bond transaction quality and make better investment decisions. The mathematical models are built to analyze the previous bond transaction price, transaction date, and trade size to produce an easy-to-understand bond trade cost benchmark for individual investors to self-measure their bond transaction effectiveness.

Team: Omokhowa Agbojo [BIOE]
Advisor(s): Tarek Zohdi [ME], Mayasari Lim, Josephine Lembong

Cell therapy manufacturing processes have a significant environmental impact even at preclinical scale. They consume large amounts of energy and water and produce plastic and chemical waste. This environmental impact is expected to scale exponentially as demand increases. We are applying the life cycle assessment tool to the hMSC manufacturing process to model how the environmental impact changes with various process parameters. Our model considers the upstream manufacturing process and includes the evaluation of environmental impact categories, uncertainty propagation, and a sensitivity analysis. This lifecycle assessment method will help engineers design more environmentally conscious processes for cell therapy production.

Team: Jacob Szymkowski [BioE], Jianshe Guo [ME], Ziang Deng [ME], Kalle Suzuki [ME], Jiahao Zhao [ME], Chunyu Jin [ME], Weibo Huang [ME]
Advisor(s): Gabriel Gomes [ME]

Onboard computer vision for robots is often limited in field of view and in localization potential, which are both vital for precise navigation. Our objective is to apply reinforcement learning to develop a framework that improves a compact robot using remote imaging systems, improving the quality of data available for navigation. This supplies the robot with information that onboard sensors cannot provide, and has the potential to reduce labor costs in a multitude of fields by expanding the range of environments robots can navigate.

Team: Yu-Chun Chen [EECS], Zhaoqing Cui [EECS], Zizheng Tai [EECS]
Advisor(s): Laurent El Ghaoui [EECS]

There is a lack of a rigorous and universal measurement framework for the robustness analysis of neural networks. Current analysis results can vary due to different measurement metrics or types of models. We seek to establish a process where all neural networks can be analyzed under the same standard, which will help us when developing and evaluating more robust models. We discover that nearly all neural networks can be transformed into implicit deep learning models, whose robustness properties are expressed through sensitivity matrices. Furthermore, adding sensitivity matrices can potentially decrease neural networks’ vulnerability. The project result is software to generate robust implicit deep learning models from different types of existing models, which can make current models in critical applications more robust against attacks with relatively low effort.

Team: Michael Khorram [EECS], Tiantian Wang [EECS]
Advisor(s): Allen Yang [EECS]

OpenARK is an open sourced augmented reality software development kit whose goal is to allow developers to rapidly prototype AR applications. Our team’s goal is to improve the accuracy of the SLAM system in OpenARK which is responsible for tracking the device’s location relative to its environment. Our primary improvements include the merging of independent maps caused by the re-initialization of the SLAM system in low-feature environments, as well as the replacement of the current image feature detector with an improved one based on new research.

Team: Jiaming Luo [ME], Winnie Lai [ME], Xiaolin Wang [ME], Xiangjiu Wu [ME]
Advisor(s): Kosa Goucher-Lambert [ME]

The product development process determines 80% of the environmental impact of a product in the design and manufacturing industry. However, the current environmental impact management strategies are time-consuming and technologically complex to use. Our team is exploring the problem and solution from the perspective of designers to develop an effective and easy-to-use interactive AI assistant. We conducted user research, market research, and data analysis to construct a decision tree of a corporation’s Design for Environment Initiatives. We will provide a well-defined project scope for future research teams to build a usable dataset for the interactive AI assistant development.

Team: Qi Deng [IEOR], Nicholas Foo [IEOR], Zihao Zhou [IEOR]
Advisor(s): Lee Fleming [IEOR], Guan-Cheng Li [IEOR], Bo Heiden, Matt Rappaport, IP CheckUps, Gregg Scharfstein, LBL, Eugene Chow, PARC

Intellectual property (IP) is an expression of human ingenuity. IP—characterized through examples like patents, trademarks, and art—builds upon existing ideas and solutions, and stretches the frontiers of existing fields. IP touches many slices of society, ranging from technology and education to agriculture. With the growth in technological innovation, the importance of maximizing the value of IP has never been more salient. Our project strives to determine unique value propositions of agriculture-based intellectual property assets through several patent analysis tools. We will work with ESpaceNet and Cipher, two tools that provide data analytics and classification modeling to inform trends within a specific space. Our team hopes to identify competitive landscapes of agriculture-based patented technologies and map out pathways to commercialization.

Team: Michael Chiu [IEOR], Tianqi Fan [IEOR], Ruila Julia Puskas-Juhasz [IEOR]
Advisor(s): Lee Fleming [IEOR], Guan-Cheng Li [IEOR], Bo Heiden, Matt Rappaport, IP CheckUps, Gregg Scharfstein, LBL, Eugene Chow, PARC

Around 30% of start-ups fail because of a non-viable business model. According to Harvard Business Review, not clearly understanding the competition and innovation are two underlying causes. Thus, it is hard for start-ups to commercialize their breakthroughs because of this lack of visibility on the competitive and innovative landscape. Our team is working on leveraging ML and NLP algorithms to predict the competition in the industry based on patent analysis, therefore enabling companies to adjust their go-to-market strategy. This algorithmic approach to a business problem will help start-ups adjust their commercialization schemes.

Team: Lance Miranda [ME], Shidian Zhang [IEOR], Yixuan Dong[IEOR]
Advisor(s): Lee Fleming [IEOR], Guan-Cheng Li [IEOR], Bo Heiden, Matt Rappaport, IP CheckUps, Gregg Scharfstein, LBL, Eugene Chow, PARC

Many breakthrough technologies developed by labs and universities reach the proof-of-concept stage, however never make it to industry. Our team is applying machine learning to patent analysis to understand a nuanced laser industry landscape and to develop a viable commercialization strategy for an ultrafast pulsed laser combination method developed by Lawrence Berkeley National Laboratory. Furthermore, our project will provide a framework that applies to other products and help move research-based technology to market and industries.

Team: Jace Sheu [IEOR], Yiheng Qi [EECS]
Advisor(s): Lee Fleming [IEOR], Guan-Cheng Li [IEOR], Bo Heiden, Matt Rappaport, IP CheckUps, Gregg Scharfstein, LBL, Eugene Chow, PARC

Companies, universities, and governments are spending exorbitant resources on innovation trend analysis in order to catch the wave of the next biggest technologies. With over three million enforced patents, there is a wealth of information with enormous potential. We are building different time-series analysis models utilizing decades of United States Patent and Trademark Office (USPTO) data to accurately predict the number of patents in future years. Analysts will be able to grasp the insights from our predictions and set meaningful strategic directions for business development.

Team: Marc Horneck [IEOR], Estella Liu [IEOR], Meina Piao [BIOE]
Advisor(s): Lee Fleming [IEOR], Guan-Cheng Li [IEOR], Bo Heiden, Matt Rappaport, IP CheckUps, Gregg Scharfstein, LBL, Eugene Chow, PARC

Scientists have estimated that our planet is expected to run out of fossil fuels by 2060. Our team is tackling this issue by creating a commercialization strategy for a newly patented sustainable biofuel molecule through machine learning and patent analysis tools. Once on the market, this new molecule will be a sustainable high energy fuel source for a variety of applications such as rockets and jets.

Team: Magdalena Ossa [ME], Yibo Luo [NE]
Advisor(s): Massimiliano Fratoni [NE], Pedro Junior Valdez [NE]

Carbon- free energy sources like solar and wind are characterized by a strong variability that, today, is compensated by firing fossil fuel plants. Nuclear Power Plants, instead, are limit to baseload generation due to legacy design choices that prevent to vary their outputs during short periods of time. New reactor designs can overcome such limitation. We use machine learning tools to first understand California’s energy market demands, and then to determine the ideal characteristics of a fleet of nuclear reactors with load following capabilities with the ultimate goal to achieve a 100% clean electric grid of nuclear and renewable sources.

Team: Xiyu He [IEOR], Lei Liang [IEOR], Abby Wang [CEE], Shae Alhusayni [IEOR]
Advisor(s): Cristobal Pais [IEOR], Max Shen [IEOR]

E-commerce companies such as Amazon, Alibaba and JD.com offer millions of products. Predicting demand for any product is complex, let alone a new product. Our project aims to accurately forecast e-commerce merchandise to help establish and upgrade product supply chain strategies, reducing inventory, and lowering stockout penalty costs. Our goal is to provide a systematic time-series machine learning solution with a high prediction accuracy. This will predict sales of new products by only using sales history data and attribute information of older products (on JD.com). This solution can also be reused for other e-commerce platforms.

Team: Diego Espinoza [BIOE]
Advisor(s): Phillip B. Messersmith [BIOE/MSE], Katerina Malollari [ME], & Kelsey DeFrates [BIOE]

Each year, millions die from vaccine-preventable illnesses, an inequality largely fueled by technical challenges surrounding the administration of vaccines with hypodermic needles. MEDiRoller, an automated microneedle device, revolutionizes vaccine and drug delivery through the elimination of trained medical personnel requirements, reduction of blood-borne disease transmission, and preservation of patient comfort during immunizations. Using a novel handheld device that controls the applied force, the MEDiRoller creates small passages over large regions of skin to allow for drug transport without trained medical professionals while mitigating sharp exposure.

Team: Carol Yan [EECS], Jinyue Zhu [EECS], Eric Wu [EECS]
Advisor(s): Kristofer Pister [EECS]

Chronic wounds, identified as wounds that do not naturally heal for over three months, affect 6.5 million people in the U.S and lack effective treatment. We aim to design a compact and affordable smart bandage to monitor various biosignals essential for wound healing and carry out real-time wound treatment. Previous smart bandages are expensive and bulky. We incorporate a Single-Chip Micro Mote (SCμM) with on-board wireless temperature sensor and a low area footprint printed-battery. With that, we can achieve a disposable smart bandage that costs $1 to identify the infection of chronic wounds.

Team: Xiao Zhang [CEE], Fengyu Zhu [IEOR], Raven Han [CEE]
Advisor(s): Max Shen [IEOR], Cristobal Pais [IEOR]

The scale of e-commerce has been growing rapidly, and this growth has brought remarkable opportunities to transform the traditional warehouses into robot-based e-commerce warehouses. E-commerce warehouses have more complex functions and thus it is of great importance to improve their efficiency to meet the needs of customers. Our project aims at finding out the best locations of picking stations so as to minimize the total travel distances of the working robots. We will mainly use integer programming methods with complex constraints that match the real warehouse conditions. We will also study how optimal picking stations locations change due to different warehouse layouts. By doing several simulations, we hope to give e-commerce companies conclusive rules on how to design an efficient warehouse.

Team: Siyi Qu [IEOR], Jiayue Tao [IEOR]
Advisor(s): Gabriel Gomes [ME]

As individuals, we are exposed to, connected to and creating open-source data every day. With the existing machine learning techniques, predictions could be made more easily than ever, but the sheer number of available models overwhelms people. In this project, the most popular machine learning tools, from support-vector machine to neuron network, will be used to make high-accuracy predictions based on numerical, text and image data. Our goal is to help people without technical background understand, apply and gain insights on machine learning models in a more efficient and effective manner, with real-life examples (e.g., Titanic, Airbnb and American Sign Language).

Team: Michael Zhang [EECS], Roham Ghotbi [EECS]
Advisor(s): Alexandre Bayen [EECS]

The city of Fremont is facing a significant challenge that is impacting the quality of life for its residents and businesses. Regional cut-through traffic, led by navigation applications, is clogging local roadways with motorists that do not live or work in Fremont. Our team is working on a dynamic traffic model that provides accurate simulations on various time periods. The resulting model can contribute to understand and mitigate congestion without decreasing road capacity in the real-world.

Team: Vianna Quenon [ME], Deshan Cai [ME], Weisa Wang [ME]
Advisor(s): Hayden Taylor [ME]

3D printing is unique in its ability to print personalized parts for dentistry, aerospace, and medical applications, yet it faces challenges when mass-produced due to the quality, timing, and scale constraints of 3D printing. To bring 3D printing to mass-manufacturing, researchers at UC Berkeley have created CAL, a faster way to 3D print with higher quality. Our team is working on a new rotating projection mechanism with multi-wavelength projection and surface roughening to improve the quality and versatility of mass production by 3D printing.

Team: Zizhao Gong [ME], Michael Cui [ME], James Cheney [ME], Chufan Guo [ME], Jiuqi Wang [ME], Yiliang Sun [ME]
Advisor(s): Allen Yang [EECS]

Autonomous driving is on the verge of changing the world, but the development needed to master the complexity of full autonomous operation is difficult to achieve with current platforms. Simulators are cost effective and accessible, but not realistic enough to test all situations. Testing with real vehicles is perfectly realistic, but expensive and complicated to implement in research. Berkeley’s Robot Open Autonomous Racing (ROAR) program addresses this need with a research platform that includes cross-compatible simulation and 1/10 scale physical environments. The Winning ROAR!! Team is making improvements to the ROAR dual environment platform as well as proving its effectiveness by developing and racing autonomous vehicle agents on it.

Team: Szu-Wei Tung [NE], Xiatong Yang [NE], Oren Stoelting [MSE]
Advisor(s): Peter Hosemann [NE]

Conventional machining of radioactive materials relies on direct contact methods that allow debris to easily contaminate the working area. This necessitates lengthy and costly downtimes for cleaning and maintenance. However, laser ablation does not require direct contact, but can instead operate through a glass barrier. Utilizing this, we have designed an enclosure equipped with a glass port for laser access, controlled airflow, and remote manipulation of the sample in order to limit debris to a smaller area. This method reduces contamination, thereby reducing cost and turnaround times while improving safety.

Team: Doris Tai [BioE], Haohong Lin [ME]
Advisor(s): John C. Anderson [BioE]

Many enzymatic reactions follow unnamed and rare mechanisms. Defining which atoms are part of the enzymatic reaction operator from those that should be excluded is still a challenge for biochemists. ROExtarctor applied atom-to-atom mapping in Java to solve the previous challenge. Our team’s work involves cleaning up the codes, testing the algorithm with examples, and generating statistics for data analysis prior to the publication. This model can benefit future biochemical research in enzymatic reactions by validating the models and automatically generating the diagrams for a better understanding of the mechanisms.

Team: Jasper Shih-Pu Lee [CEE], Han Wang [CEE], Kim Muy Ly [CEE]
Advisor(s): Alexander Skabardonis [CEE], Offer Grembek [CEE]

Every year, over 1 million people are killed on roadways around the world. Our team has created an application that allows users to voice their concerns about infrastructure designs that make them feel unsafe. Through the data received by our mobile application, we can identify which road features make road users feel unsafe during their travels, so that local municipalities can proactively identify and improve dangerous components of their infrastructure.

Team: Steven Johannemann [IEOR], Zoe Ouyang [IEOR], Chloe Zhang [CEE]
Advisor(s): Anil Aswani [IEOR]

Acquiring a new customer can cost five times more than retaining an existing customer, so it is crucial for businesses to stay attuned to the needs of existing customers, learn from their feedback and make decisions that align with their expectations. Our team has created a sentiment analysis algorithm that is capable of categorizing positive and negative customer reviews and identifying frequent topics from complaints. This algorithm offers valuable business insights while requiring minimal technical knowledge to operate and understand, making it helpful for all kinds of businesses, especially smaller ones that are unable to hire technical personnel.

Team: James Burke [IEOR], Xuerui Song [IEOR], Ishan Tikku [IEOR], Alejandro Gordo Cuadrado [ME], Andrew Min [CEE]
Advisor(s): Tarek Zohdi [ME]

While most understand the danger of fighting fires on the ground, more than one in four deaths among wildland firefighters result from an aviation-related incident. SimDrop uses machine learning to find the optimal altitude, airspeed, flight path and retardant flow rate for high-wind conditions, which are notoriously unsafe and detrimental to drop accuracy. Our solution incorporates a genetic algorithm and a simulated environment to test and evaluate the effect of retardant on a dynamic fire environment with various wind intensities, thus improving aerial firefighter safety and efficacy.

Team: Rebecca Sung [ME], James Liao [ME], Henry Fong [ME]
Advisor(s): Reza Alam [ME], Alexandre Immas [ME]

81% of the ocean remains unexplored, due to the ineffectiveness of traditional, wireless technology for underwater communication. Current methods mostly utilize acoustic communication, which is hampered by its low speed and limited bandwidth. Our team is laying the foundation for a new wireless optical communication method which localizes and transfers data using a communication link formed by a swarm of unmanned underwater vehicles (UUVs). The advantage of this new approach is twofold: it improves communication speed and bandwidth, and also allows longer range and depth for ocean exploration than is currently possible.

Team: Yijie Huang [MSE]
Advisor(s): Junqiao Wu [MSE]

This project is based on the insight that elevating road temperature by modifying the emissivity pattern of concrete roads could impede the formation of ice and snow in an environmentally friendly way. Heat radiation contributes to large portion of heat lost. Heterogenous materials that are composed of metal nanoparticles and concrete can preserve large amount of heat from infrared emission. We used the binary composites model for our metamaterials to calculate materials’ effective permittivity, which depends on the inclusion filling factor and depolarization factors.

Team: Rita Chen [ME], Michael Kotur [ME], Danny Ma [ME], Ian Miller [ME]
Advisor(s): Reza Alam [ME], Michael Kelly [ME]

Coastal communities affected by natural disasters lack affordable and reliable access to drinking water. We aim to create a portable device that captures wave energy to drive seawater desalination and provide freshwater to communities without the need for external power or materials. The portable device is composed of a pumping mechanism that brings seawater to a reverse osmosis system to generate drinking water. Additionally, the device can be assembled without assistance from a professional, resulting in user independence for gathering freshwater.

Team: Alexa Gomberg [BIOE], Kate Zhao [BIOE], Natalia Matti [BIOE], Helen Chen [BIOE]
Advisor(s): Emily Carvalho [BIOE], Cameron Morley [BIOE], Sanjay Kumar [BIOE]

Inefficient delivery of drugs, in combination with the 11-14 year process of drug development, contribute to why many diseases are still difficult to treat despite decades of research. Our capstone team has conducted a literature review to summarize recent advances in drug and cell delivery with hyaluronic acid hydrogels. Access to such information can facilitate the improvement of existing disease treatments efficacy by allowing drug developers to easily explore and compare HA delivery platforms comprising different material structures and modifications.

Team: Zhe Ding [ME], Gibreel El-Halabi [ME], Hector Gomez [ME], Shuo Zhao [ME]
Advisor(s): Koushil Sreenath [ME], Zhongyu Li [ME], Bike Zhang [ME]

The advancements in today’s humanoid robots, capable of walking through complex terrains, jumping, and climbing up stairs make them viable to assist first responders in addressing less than ideal situations. The robot, Digit, was not initially designed to assist first responders, but our team is working to bridge that gap by advancing the science of bipedal robots, done by designing and developing a Model Predictive cascade control method. The method uses this optimization method to find an ideal joint configuration (currently not mature for humanoid robotics), followed by a Whole Body Impulse Control to generate actuation.