In my 3rd year of university I led my first QMIND project. We worked on a system to apply Alpha Zero, the state-of-the-art chess and Go bot, to currency trading. The setup for this is pretty awesome, and I want to write about it in the future, but you can check out an explanation here. The idea was to use the training techniques from Alpha Zero (and a graph neural network) to find shortest-paths in a graph. If you pre-process your currency graphs correctly, the shortest path may correspond to an arbitrage opportunity. At a glance we...

• reimplemented AlphaGo from scratch in PyTorch and trained it for 36 hrs on random graphs
• hacked together a UI to show the model processing live cryptocurrency exchange graphs
• placed top 8 at CUCAI and presented our work at the Amazon office in Toronto!