Numerical Methods for Solving Optimization Problems
Have you ever thought of how computers solve optimization problems? In this project, I wrote a Wikipedia/Scholarpedia hybrid style article on solving optimization problems with numerical methods. I divided most of the report into two sections: constrained optimization and unconstrained optimization. I discussed optimality conditions and methods that are essential to solving optimization problems.
Senior Design Capstone: Convoy Shipment Process Improvement
Convoy is a digital freight broker that aims to make freight more efficient and environmentally sustainable. In my industrial engineering senior design capstone project, my team partnered with Convoy to improve their processes in scheduling and handling loads for shippers, with the goal of mitigating conflicting appointment times (CATs). We analyzed Convoy's internal processes and conducted data analysis to determine the root causes of CATs. After we determined the problems, we created the following deliverables for Convoy:
Machine learning model that predicts when CATs may occur in future shipments.
A standard operating procedure to audit shipments that are flagged to be at risk for a CAT by the machine learning model.
Recommendations on improving visibility of new load confirmation details when updates to shipments occur.
Recommendations on additional data to collect regarding load confirmations.
Recommendations on changes to a UI platform used by Convoy's operators.
We saved Convoy approximately $1,000,000 per year, 4.3 driving hours per shipment, and 190,000 miles of driving per year.
From all of our hard work, my team was awarded as 1 of 3 finalist teams out of 28 total industrial engineering teams for the Spring 2022 semester! We were given the honor of presenting our project one last time in the finalists presentation. Thank you to Convoy for an incredible collaboration and to my team's faculty advisor Prof. Leon McGinnis!
My team at the Spring 2022 Georgia Tech Capstone Design Expo
Machine Learning for Wildfire Susceptibility Mapping
In an era of climate change, wildfires have become a focal point of study - particularly for the West Coast. Trends for recent years have not shown much change in the total amount of fires, but they have shown an increase in the amount of land burned by these fires. The implications can be severe: deforestation, homelessness, and increased carbon emissions, among others. In this project, my team created a machine learning model to classify and predict susceptibility of states in the Contiguous U.S. to wildfires in 2020. We collected and cleaned data on wildfire counts, natural phenomena, and state demographics over time. For machine learning, the team first developed two methods to handle feature selection in the dataset:
Correlation matrix: Gives a quantitative sense on features correlated well with wildfire susceptibility.
Principal Componenet Analysis: Unsupervised machine learning technique to shrink the number of dimensions in the model.
Afterwards, the team developed three methods to predict wildfire susceptibility:
Naive Bayes: Supervised machine learning technique to classify wildfire susceptibility into certain categories based on magnitude of susceptibility likelihood.
Linear Regression: Supervised machine learning technique to predict the number of wildfires in the coming year.
Logistic Regression: Supervised machine learning technique to predict the number of wildfires in the coming year.
Investigating the Effects of Ramp Metering on Traffic Flow in Complex Traffic Systems
Atlanta is labeled as one of the most congested cities in the United States. To alleviate congestion, Atlanta installed ramp meters in designated locations along interstates. In this project, my team designed a discrete-based simulation of the effects of ramp metering on traffic flow. We undertook an object oriented design to monitor traffic simulation and vehicle ramping strategies. We utilized a strategy called ALINEA to determine ramp metering rate. We also utilized a modified ALINEA strategy, which takes in the ramp queue length to determine ramp metering rate. The case study was on the I-75/I-285 interchange using real world data on traffic volume. The results show that ramp metering does improve traffic flow. However, the ramp metering rate should be monitored and adjusted according to the queue length in order to avoid an unbounded ramp queue.
Case Study Area and Traffic Flow Illustration for Project
The Minimum Vertex Cover Problem is an NP-Complete problem with applications in many different fields. In this project, my team designed four different algorithms to solve the minimum vertex cover problem:
Branch and Bound Algorithm
Approximation Algorithm
Stochastic Local Search Algorithm
Simulated Annealing Local Search Algorithm
These algorithms were tested on real world datasets from the 10th DIMACS challenge. We undertook an empirical evaluation to assess effectiveness of each algorithm on the datasets.
Two Minimum Vertex Cover Concept Examples Image Source: Wikipedia
