Bill Hu

Chrome Dino Mini Game

ShanghaiTech University, Computer Architecture I (CS110)

In this project, we implemented a Chrome Dino Runner Game on the Longan Nano, a risc-v development singlechip. We utilized PlatformIO to interact with the development board and support cross-language programming. We implemented most of the framework by C, but some of the auxiliary functions are implemented by RISC-V assembly. The functions can be invoked bidirectionally.

The game we implemented features:

• A simple startup menu
• A difficulty-selection panel
• A live score board, which increases as T-Rex running
• Randomly generated cactus and pterosaurs acting as obstacles
• Two buttons to control the T-Rex to jump or creep to avoid obstacles
• A invincible mode in which T-Rex will not be killed by obstacles, activated periodically by score.

Close Project

Multi-armed Bandit Analysis

ShanghaiTech University, Probability and Statistics (SI140)

This project is about Multi-armed Bandit, a classical problem in reinforcement learning. The project is generally divided into the following three parts:

• Simulation. We have to simulate the multi-armed bandit via different algorithms under the set of given parameters and show the output, i.e., the performance of the bandit.
• Analysis. We have to analyze the impact of each hyperparameter of each algorithm. Additionally, we have to analyze an important problem in the multi-armed bandit program, which is named the exploration-exploitation trade-off.
• Design. We have to handle some variants of the classical scenario, including dependency and constraint, and design some algorithms by ourselves to address the novel problem variants.

You can refer to the outline to learn about the background and requirements of this project and refer to the report to look through my project report.

Close Project

AI Aircraft Warfare Game

ShanghaiTech University, Artificial Intelligence (CS181)

In this project, we implemented an AI-agent for playing the Aircraft Warfare game, using Approximate Q-learning and Deep Q-learning.

Aircraft Warfare is a classical game in which the player controls an aircraft to shoot down the enemy aircrafts. The player can move the aircraft left or right and shoot bullets to attack the enemy aircrafts. The player will lose the game when hit by the enemy aircrafts.

In this project, we implemented an AI-agent for playing the game with Approximate Q-learning method and Deep Q-learning network (DQN) method. For Approximate Q-learning, we extracted four most useful features and trained our agent with an online learning method. For DQN, we utilized a convolutional neural network (CNN) which takes the screenshot as input and outputs the expected converged sum of the discounted reward of actions given the current state.

Close Project

Ray Tracing NURBS Surfaces

ShanghaiTech University, Computer Graphics I (CS171)

A tessellation-free NURBS surfaces ray-tracing renderer.

As a non-analytical surface, high-quality-and-efficiency NURBS surface rendering is a very challenging topic in Computer Graphics. Traditional triangulation (tessellation) method requires a huge amount of calculation and memory to generate and store meshes respectively. To solve this problem, in this project, we implemented a path tracer with global illumination to directly render untrimmed NURBS surfaces without tessellating the surface.

Following the methods proposed by Martin et al, we successfully implemented the renderer with more than 3300 C++ baseline. In details, we applied surface refinement for better rendering quality and efficiency, and constructed a BVH to maintain the surface sub-patches and accelerate the ray-patch intersection process.

Close Project

RV32I-RVC Bi-directional Translator

ShanghaiTech University, Computer Architecture I (CS110)

In this project, we wrote a translator that translates RISC-V instructions to RISC-V compressed instructions and vice versa. The RISC-V standard compressed instruction set extension (called RVC for shorthand) offers 16-bit versions of common 32-bit RISC-V instructions, thus reducing the binary machine code size.

For more details please refer to the course page: First part and Second part.

Basically, it is not a hard job to understand but it takes us a long time to understand the RVC encoding rules and debug the C89 and risc-v assembly code. The following picture is my draft paper about it. The result is satisfying: the translator can parse the machine code, adjust jump offsets, with no memory leak.

Close Project

Machine Learning for Recommender Systems

ShanghaiTech University, Intro to Machine Learning (CS182)

Recommender systems can help predict users' unspecified needs by recommending potentially interesting content for users. It has been widely applied in many fields,such as online shopping.

In this project, we thoroughly researched the state-of-the-art algorithms in recommender systems domain. We re-implemented these algorithms, designed comparison experiments, and analyzed the performance, efficiency and usage scenarios. The algorithms we re-implemented are:

Simple Algorithm for Recommendation (SAR): The core idea of SAR is to calculate the similarity between items and users by multiplying the similarity matrix (item-item) and affinity matrix (user-item). The result are then sorted by score for each user to get the recommendation list.

Neural Collaborative Filtering (NCF): NCF is a simple and generic nerual network architecture, which shares the idea of Voting. We explored an NCF algorithm which is a fusion of Generalized Matrix Factorization (GMF) and Multi-Layer Perceptron (MLP).

Wide & Deep Neural Network: The Wide&Deep model is a joint of linear model (wide) and neural network (deep). The 'wide' part is used to memorize the user's historical behavior, while the 'deep' part can learn user's interest from the historical behavior. In this way, the model can achieve both memorization and generalization.

Close Project