Coming soon…

Every year Machine Learning course (CIS520) hosts a project competition. The challenge in 2016 was to classify the given tweets into 2 categories, namely “happy” and “sad.” Two data sets were provided, one of which is labeled and the other unlabeled. Each set consisting of 4500 observations included the top 10000 words observed in the collection, raw tweets, word counts for each tweet, downsized images linked to the URL in the tweets, and some CNN features and color mappings related to images. This submission won the **3rd prize** in the competition out of 60 teams. The report below contains the summary of 1) feature extraction and feature manipulation techniques for text classification which includes dimensionality reduction by removing missing features, Principal Component Analysis or Stemming, choosing top words on the basis of information gain, and addition of bigrams for making features more informative, and 2) different supervised and semisupervised classification algorithms based on Logistic Regression, Support Vector Machines, Naive Bayes, Mixture Models and Ensemble Methods and the results obtained for the above problem.

- A. Jain, K. Jang. Classification of Tweets using Supervised and Semisupervised Learning. CIS520 Machine Learning Competition, University of Pennsylvania, 2016. [pdf]

This work was done as a part of Advanced Robotics course (MEAM620) at UPenn. We implement a controller and a trajectory planner on a Nano+ quadrotor and test it in GRASP Lab at UPenn. Two PID controllers, one for position and the other for attitude are used. A* algorithm is used to avoid obstacles while planning the trajectories. Two types of trajectory generators are implemented, namely cubic and quintic splines for minimum acceleration and minimum jerk, respectively, depending upon the the nature of defined path.

- A. Jain, Y. Oquendo, M. Gilbert, E. Young. Implementation of Planner and Controller. MEAM620 Advanced Robotics, University of Pennsylvania, 2016. [pdf]

MLE+ is an open-source tool for energy efficient building automation design, co-simulation and analysis developed at the Real-time and Embedded Systems Lab at UPenn. It provides a Matlab interface to run the energy simulator EnergyPlus. Its graphical front-end can be used for system identification and designing advanced control strategies, in which the building simulation is carried out by EnergyPlus while the controllers are implemented in Matlab or Simulink. After a co-simulation, the output data from EnergyPlus can be aggregated, analyzed and visualized in Matlab.

Benefits of an an electrically assisted turbocharger in a conventional vehicle are evident in fuel consumption as well as acceleration performance. We investigate its suitability in a hybrid electric vehicle (HEV) with a turbocharged engine. This system is based on the design of Formula Hybrid technology with two electric machines, a traction motor and a boost motor coupled to the shaft of the turbocharger, and offers an additional control variable in the energy management problem i.e. the amount of electrical boost (e-boost) to reduce the turbolag. The task of an optimal controller now becomes manifold: deciding the torque split between the engine and the traction motor, the power of the boost motor and the gear number. A quasi-static model of a parallel HEV with a turbocharged engine is derived and a method to model turbolag based on a predefined map for the permissible engine torque is proposed. Dynamic Programming is used to solve the optimal control problem. The circumstances in which it is advantageous to use the boost motor are discussed. Further, the influence of powertrain components’ size on the control strategy is analyzed, specifically in the problem of maximizing acceleration performance.

This thesis was done as a part of research on Energy Management of Hybrid Electric Vehicles at ETH Zurich and was sponsored by Daimler AG.

- A. Jain, T. Nüesch, C. Naegele, P. M. Lassus, C. H. Onder. Modeling and Control of a Hybrid Electric Vehicle with an Electrically Assisted Turbocharger. In IEEE Transactions on Vehicular Technology, 2016. [pdf]
- Master Thesis: Optimal Control of a Hybrid Electric Vehicle with an Electrically Assisted Turbocharger, ETH Zurich, Switzerland, 2014. [pdf]

Meta-model based optimization is an efficient tool to optimize computationally expensive and noisy models, like finite element models (FEM) for example. Consider the geometric design of a thermal insulator. Due to computational complexity, it is not possible to simulate thousands of finite element models with different geometric designs to come up with an optimal design. Further, meshing also introduces noise in the simulations. Shape optimization with meta-modeling overcomes these limitations. We uniformly sample the design parameters and learn a regression model that predicts the desired metric of performance, for example heat loss in this case. We then perform optimization on the trained models or meta-models to choose the optimal design. In this project, we develop a proof-of-concept for this approach. FEMs are constructed using FAESOR toolbox for MATLAB, and statistical meta-models using different ordinal and standard regression approaches.

- J. Poland, A. Jain, K. So. Ordinal Regression for Meta-Modeling in Optimization. Technical Report, ABB Corporate Research, Switzerland, 2014.

This project presents a study on the design of linear model predictive control (MPC) for wind turbines, with a focus on the controller’s tuning tradeoffs. A continuously linearized MPC approach is described and applied to control a 3-bladed, horizontal axis, variable speed wind turbine. The tuning involves a multiobjective cost function so that the performance can be optimized with respect to five defined measures: power variation, pitch usage, tower displacement, drivetrain twist and frequency of violating the nominal power limit. A tuning approach based on the computation of sensitivity tables is proposed and tested via numerical simulations using a nonlinear turbine model. We further compare the performance of the MPC controller with those of a conventional one.

- A. Jain, G. Schildbach, L. Fagiano, M. Morari. On the design and tuning of linear model predictive control for wind turbines. Renewable Energy, 80, 664-673, 2015. [pdf]

A brain machine interface (BMI) can be used to decode brain signals and control a robotic arm. We discuss its application as a prosthetic device, especially for spinal cord injury victims, suffering from paralysis. Neural data for experimentation is recorded from intra-cortical electrodes implanted in a macaque monkey at National Brain Research Center (NBRC). Brain signal decoding is achieved using the population vector algorithm (PVA). A delta manipulator with compliant linkages has been designed integrated with the BMI to perform a standard left-right instructed-delay centre-out reach task. An automatic raisin dispensing mechanism has also been designed and used to automate the complete experiment.

A working prototype of BMI can be seen in this video.