The following is a list of current projects between TRIPODS investigators and their collaborators. If you are a researcher and are interested in collaborating with a TRIPODS investigator, please contact us. Contact information can be found on each investigator’s website, links for which can be found on the People page.
Scaling up derivative-free optimization methods for reinforcement learning and robotics
This project is supported by a Google Faculty Award and in collaboration with Google Brain. The simulation-based reinforcement learning problem that is considered in this project has generated interest in purely black-box methods for reinforcement learning. However, the scale of the problem is much larger than that of traditional simulation-based optimization problems. We explore several parallel and sequential methods for solving problems in this setting with specific application to robotics.
Optimization based protein alignment and protein function prediction
In this joint project with a bioinformatics expert from the CSE Department at Lehigh, we explore similarity in specificity of proteins based on their optimal alignment. This alignment is discovered through optimization of a noisy black-box estimate of the volume of the intersection of the proteins. This information can further be used to predict specificity of the proteins.
Convergence rates of stochastic optimization methods
This project considers numerical methods for solving very broad classes of optimization problems. We analyze a variety of optimization methods—such as stochastic gradient, line search, and trust region methods—under various assumptions on the accuracy of computed function and gradient estimates. The goal is to identify conditions under which stochastic problems can be solved by methods whose convergence behavior matches that of the methods in a deterministic setting.
Optimal classification trees
This project is in collaboration with researchers at IBM T.J. Watson Research Center. We are developing a novel mixed integer programming formulation to construct optimal decision trees of a pre-specified size. This formulation takes the special structure of categorical features into account and allows combinatorial decisions (based on subsets of values of features) to be made at each node. Interestingly, we have found that very good accuracy can be achieved with small trees using moderately-sized training sets.
Efficiently training large ML models on huge datasets
In this project, we explore new algorithms for big data applications that have the ability to scale on large clusters. In particular, we design new methods for training deep neural networks and recurrent neural networks in the context of prediction models and reinforcement learning. The goal is to design methods that reduce communication costs and, hence, work much faster in distributed computing environments.
Automated scheduling of production and delivery of perishable items
This work is in collaboration with researchers at Siemens Corporation, Corporate Technology (SC CT). We address the problem of scheduling the production and delivery of perishable items, i.e., items that have a very short life-span, such as radiopharmaceuticals used during PET scans. In this setting, it is critical to accurately forecast the demand for the product and account for this demand in the scheduling. Other factors need to be considered as well, such as the effects of weather on travel times. We utilize reinforcement learning to train predictive models using a massive amount of historical data to design the production and delivery schedule.
Active metric learning for supervised classification
This project is a part of a large collaboration with ExxonMobil Research. We aim to learn a metric by which one may identify outliers versus essential samples in a dataset. This technique helps in selecting the next sample in an online setting and in reducing the size of datasets in a batch setting to achieve faster training.
Efficiently solving large-scale alternating current optimal power flow (ACOPF) problems
ACOPF problems are constrained polynomial optimization problems that need to be solved repeatedly in order to adjust the voltages and phases in a power network. We are working with IBM research in Ireland on various vital aspects of ACOPF; e.g., we consider the issue of how to efficiently handle time-varying loads in a power network.
Next generation structural health monitoring
The existing structural health monitoring (SHM) paradigm relies on data that is expensive to generate by existing processes. In this project, we aim to utilize data from non-traditional sources such as mobile sensors and networks of cheap sensors. Such data is then fed into a deep learning platform to create new information about the state and condition of structural systems and infrastructure.
Using constrained optimization techniques to solve data science problems
This project focuses on exploiting the wealth of methods for solving constrained optimization problems in order to solve data science problems. For example, in collaboration with researchers at Columbia University, we are developing new alternating direction algorithms for solving optimization problems with multiaffine constraints, which has applications in dictionary learning and phase retrieval.
Applications of machine learning in science and engineering
This header covers various collaborations between TRIPODS investigators and other faculty members at Lehigh University. One such project focuses on developing and applying machine learning methods to find correlations in multi-channel data for nano-materials (specifically, Raman maps of graphene samples), and to detect hidden physical variables (such as strain, interlayer coupling, and doping). Another project aims at quantile regression using SVMs and neural networks for accurate wind power prediction.