MCTS-T(+)

Code for the paper Monte Carlo Tree Search for Asymmetric Trees by Thomas M. Moerland, Joost Broekens, Aske Plaat and Catholijn M. Jonker.

Prerequisites

1. Install recent versions of:

• Python 3
• Tensorflow
• Numpy
• Matplotlib

2. Clone this repository:

git clone https://github.com/tmoer/mcts-t.git

Syntax

You can run a new experiment from the agent.py function. Hyperparameters can be parsed through the --hp option. Default hyperparameters are listed in mcts-t+/hps.py. For example, to start a default experiment on CartPole-v0:

cd mcts-t+
python3 agent.py --hp game=CartPole-v0

Reproducing Paper Results

The results of the paper can be reproduced by:

cd mcts-t+
bash jobs/paper_jobs.sh

This automatically loop over the necessary hyperparameters. Running it will take quite long on a regular laptop though. You can submitted the runs to a SLURM cluster via

bash jobs/paper_jobs_slurm.sh

Visualization of Results

Subsequently, you can visualize the output with

cd mcts-t+
python3 visualize.py --home --plot_type mean --game your_game 

for some your_game of your choice.

Citation

@proceedings{moerland2018monte,
	author = "Moerland, Thomas M and Broekens, Joost and Plaat, Aske and Jonker, Catholijn M",
	journal = "arXiv preprint arXiv:1805.09218",
	title = "{Monte Carlo Tree Search for Asymmetric Trees}",
	year = "2018"
}