WIP: to-do: sequence data training
Inspired by Amortized Planning with Large-Scale Transformers: A Case Study on Chess arXiv:2402.04494
Using the latest transformer models, the team was able to approach the leading chess AI performance using Monte Carlo Tree Search (MCTS)
This project uses generated data instead of actual game play data for training. The recommended training data sizes were provided as:
10 - 20k : prototype 100 - 200k: practical use 500k - 1m: production optimal
Features being included: Game context, player tendencies, temporal dynamics, opponent modeling, etc.