I use Deep Reinforcement Learning (DRL) to train computer agents to walk, hop, or even play an Atari 2000 game!
The best performing agent after 4 hours of training. Yeah, it's difficult to learn to walk like humans.
Our trained agent (the green player) learned to win consistently in Pong!
My more recent work focuses on scaling DRL to Real-time Strategy (RTS) games, in collaboration with my advisor Santiago Ontañón.
Let's safely land our shuttle to the lunar surface. Gotta be steady and precise!
A simulated RTS game. The white, green block and circle are bases, resources and workers. The purpose of the game is to gather resources, build an army and destroy enemy forces.
A trained agent learning to hop with a single leg
You can easily use my DRL library CleanRL to train the agents to play all of the games above.
Feel free to get in touch with me at [email protected]. Btw I am looking for an internship position for doing DRL in 2020/2021. Please contact me if you are interested 🎉
🔧 My Projects
📖 My Publications
Huang, S., Ontañón, S., “Comparing Observation and Action Representations for Reinforcement Learning in µRTS”, AIIDE Workshop on Artificial Intelligence for Strategy Games, October 2019
Huang, S., Grethlein, D., “Generating Interpretable Class Model Visualizations for CNNs with varying Dilation Factors”, preprint, June 2019
Huang, S., Healy, C., “StreetTraffic: a Library for Traffic Flow Data Collection and Analysis”, ACMSE 2018 Conference, March 2018