Woojae Kim

This work (see write-up here) presents a Deep Q-network (DQN) with experience replay to solve Open AI Gym’s Lunar Lander environment. The core algorithm replicates that of the legendary Atari paper by Minh et al. in 2013. The trained DQN successfully controls the space fleet in a continuous state-space, enabling safe landings. The DQN utilizes three fully connected layers with rectifier nonlinearities to approximate the optimal action-value function. Experience replay and bootstrapping method through network duplication were utilized to improve the network’s performance. Extensive grid search was performed to fine-tune the network’s hyperparameters and study their effects on the agent’s performance

This work (see write-up here) explores data-driven machine learning techniques; five classification algorithms—decision tree, neural networks, AdaBoosted decision trees, support vector machines, k-nearest neighbors—were implemented on two different datasets: adult income dataset and wine quality dataset. After some data preprocessing, the algorithms were trained on both datasets through k-fold (k=10) cross-validation method. During the training, learning curves and complexity curves were plotted. Then the trained models were tested on the unseen, held-out testing sets (20% of the entire datasets), and the resulting performances were analyzed utilizing various metrics, such as precision, recall and f1-score.

I have teamed up with my former classmates at UIUC to design and prototype an air-purifying robot, Yuikhan (유익한 로봇). Yuikhan is a smarter version of robotic vacuum-for air quality control. It autonomously creates map of the living space using LiDAR and SLAM technology. After recogizing your space, it then purifies the space based on indoor air quality map that it has generated. We have built the robot from scratch using a 3D-printer and off-the-shelf components. We then integrated the hardware to and the cleaning algorithm through Robotic Operating System (ROS). We have a working protype of the device which syncs with a mobile application for direct control. See our landing page for more information.