Group members:
- Alex Kim
This is the searchless chess project, inspired by this recent paper from Google DeepMind.
A key bottleneck for actually training a good chess playing transformers is that we need scale. Millions of parameters at a minimum, which I haven't been able to run on my machine due to hardware constraints.
The src file contains the source code for the searchless_chess.src library written by Google DeepMind researchers. Inspired by their work, I present compare_policies.ipynb as the executable for this project. It is designed to run in a standard jupyter env. It contains code that examines three different policies 'action value', 'behavioral cloning', and 'state value'. It also compares the 'action value' policy at three differnet engine temperatures.
In this repo is also the pdf of the original paper and my presentation slides.
Requirements (may not need all):
absl-py apache-beam chess chex dm-haiku grain-nightly jax jaxtyping jupyter numpy optax orbax-checkpoint pandas scipy typing-extensions