Gallery (outdated)

Back in 2016, we were the first team to show that AI (BiCNet) could master human-level micro-management skills in playing StarCraft battles games.

image52-7495.gif
image50-7493.gif
image49-7492.gif
image51-7494.gif

Our team of Mean-Field Q-learners savage DQNs.

Picture1.png
image140-8920.gif

Multiagent techniques could efficiently dispatch Uber orders such that the supply demand gap in the CBD area can be significantly decreased throughout the whole day. 

ezgif.com-gif-maker-2-15951.gif

Revenue-oriented order dispatch

ezgif.com-gif-maker-3-15954.gif

Multi-agent powered order dispatch

We can use the multi-agent reinforcement learning techniques to mimic the spontaneous Mexican wave in the stadium.

pasted-image-8697.jpg
image120-8706.gif

Our multi-agent policy evaluation techniques can find the best combination of joint strategy profile for high-way driving, meanwhile finding Nash equilibrium on two-player game is known to be PPAD-complete (i.e. very hard).

image40-6672.gif
image41-6673.gif