Language Technologies Institute, Carnegie Mellon University
We develop LittleMoochi, an AI-based virtual pet that aims to help children form healthy eating habits. Visit moochi website here!
Wiki Discourse Segmentation
I developed a feature-based discourse segmentation tool for 11-724 Human Language for AI course project. WinDiff score was ranked top 3 in class!
Sparse Reinforcement Learning via Non-convex Feature Selection.
We proposed a nonconvex feature selection method for Reinforcement Learning to find a compact feature subset. Experiments on Chain Walk and Inventory Control exmaples showed lower approximation error and variance with higher sparsity.
Beat You in Rock-Paper Scissors!
Designed an intelligent game agent to play RGP game with human player. The game agent learns human strategies via an online reinforcement learning method, and dynamically changes its strategy against the human player.
DeepLearn Image Classifier
Implemented a Deep Belief Network (DBN) Classifier on images. The dataset contains 1000 gray scale images in 13 classes. Did Experiment on DBNs with different numbers of layers and nodes. We also implemented Auto Encoder Network and Convolutional Neural Network.
Twitter Sentiment/Weather Analysis
Designed an regression model to classify weather and weather-related sentiments in tweets. Proposed a novel feature function to enhance word frequncy. The problem comes from a Kaggle Competition Partly Sunny with a Chance of Hashtags.
Bluetooth Bicycle Lock
In Peking University, students like riding bicycles. As the bicycles are so crowded, it often takes students a lot of time to find their bicycles. We designed a Bicycle Lock System that enables users to lock and find their bicycles with their smart phone.
Database Tuning: 3NF/BCNF
Investigated the performance differences between 3NF and BCNF in database design. Designed a pattern that satisfies 3NF, but conflicts BCNF.Designed a program to generate test data based on that pattern. Did experiments on Microsoft SQL Server to compare the space cost, updating cost and query cost of the two normal forms. We also investigated the query execution plan to justify the performance differences.
Microblog Analysis Tool
Developed a microblog application to analyze social netwroks in Sina Weibo. It has 3 features:
Eisner Dependency Parser
Based on Jason M. Eisner's Bigram lexical affinities model, I developed an English dependency parser with user-friendly GUI.
Pipelined CPU Simulator
Designed and implemented a RISC CPU simulator supporting 32-bit instruction set. It is a 5-staged pipelined CPU. We compared its performance with none-pipelined CPU simulator. The experiment showed that ours had great advantage. We also designed and implemented cache and memory for the CPU simulator, so that we could run assembly programs such as Bubble Sort on the simulator.
Vertical Search Clustering
Developed a tool to divide search engine results into clusters. The tool contains a web crawler and a clustering Machine. We investigated several clustering algorithms including K-Means and hierarchical clustering algorithms.
GomoKu Game AI
Developed a Gomoku (5 in a row) Game Platform. Developed an AI for the game. The AI is on the top 15 among 300 AIs in a competition held by Computer Science Department, Peking University.