Won 1st place ($10k) out of 80 teams in the annual MIT Poker AI Tournament. Used VR-MCCFR+ to approximate nash equilibrium (with extra exploitative refinements) of the variant. Had full training, evaluation, and visualization infrastructure with heavy optimizations. Our final strategy was trained for 3 trillion hands over 5 days on 128 cores with a final winrate of 96% (half the losses as 2nd place).
What if you could cheat with a teammate in Leduc Poker? The strategy actually becomes exponentially more complex! Using a team-belief DAG with a game theoretical compression scheme, achieved best-in-class exploitability and an undefeated record with the highest final PnL.
I aim to solve Pokemon. So far, there have been no successful attempts at finding the nash equilibrium of competitive singles.
Placed 5th/89 in the Bounty variant of the annual MIT Poker AI Tournament. MCCFR from scratch in C++.
Created a brand new package and CI/CD pipeline for a new Kindle feature. UI was done in Jetpack Compose (super cool). Should be released soon 🤫🤫
Done at the Plasma Science Fusion Center. Learned some plasma physics and then extended + trained an existing active learning pipeline on EPED. The goal of active learning is to use less data to train a model.