Python

This project compares the performance of a Q-learning based reinforcement learning approach with the basic strategy advised by online guides for learning to play Blackjack. The results show that the Q-learning algorithm significantly outperformed the basic strategy, achieving a higher win rate and average return over a large number of simulated games. The findings suggest that Q-learning can be an effective approach to learning to play Blackjack, offering a better strategy than the basic strategy recommended by online guides.

Submission is in progress

Submission is in progress

3D human pose tracking is difficult. Recent works in the computer vision space have made major strides in estimating 3D human motion from monocular video. The core aspect of our project is the question：can motions captured from video be used in games? To answer this question, we built a VR game in Unity that simultaneously utilizes both video-distilled motions and traditional motion capture motions. This allows us to evaluate the accuracy and usability of these motions in their native 3D environment.

Humans posess a vast sensorimotor knowledge base that must be utilized to accelerate robot policy learning. While existing methods of demonstration generally ask humans to directly manipulate the robot’s state, this work utilizes a more natural demonstration capture platform (i.e. mouse and keyboard) to train deep learning networks on a motion planning task. Through qualitative and quantitative analysis, we demonstrate that such methods can successfully supervise imitation learning policies.

By learning the common structure between all LDR-to-HDR conversion tasks, our model is able to adapt it’s predictions given extra exposures of a scene. This novel approach reframes LDR-to-HDR conversion as a meta-learning problem.

This work proposes a regularization technique for meta-learning models that gives the model designer more control over the information flow during meta-training. Our method consists of a regularization function that is constructed by maximizing the distance between task-summary statistics, in the case of black-box models and task specific network parameters in the case of optimization based models during meta-training. Our proposed regularization function shows an accuracy boost of ∼ 36% on the Omniglot dataset for 5-way, 1-shot classification using black-box method and for 20-way, 1-shot classification problem using optimization-based method.

This final project explores the feasibility of using a GAN based style-transfer approach to generating semantic segmentation masks for autonomous driving application. Specifically, CycleGAN and pix2pix are benchmarked against a more traditional U-Net architecture built off of Resnet-18. The project uses a subset of the image data from the ApolloScape open dataset.

Millimeter-wave radar has the potential to offer a weather-agnostic sensing solution to pedestrian classification. To this end, I applied a CNN architecture to micro-doppler spectrogram data simulated through Matlab in an attempt to distinguish between a pedestrian radar motion spectrogram and a bicycle radar motion spectrogram.