Based on bootleg AlphaZero.
This repository implements GRAB0 with two approaches - GBTS (Gradient Bandit Tree Search), which is MCTS but with Gradient Bandits, and SGBS (Stacked Gradient Bandit Search) with a stack of Gradient Bandits.
Below is SGBS in a nutshell:
Experimental results for GRAB0 can be found in results.md. There are a few more that are currently not pushed yet, more information will follow here.
Below is the result of training 33 runs on Cartpole using the GRAB0 approach in combination with SGBS.
Installation
Run ./setup.sh. The src folder will contain a new binary GRAB0. You can adjust experiments in simulations.json.