This repository is a Python implementation of the MuZero algorithm. It is based upon the pre-print paper and the pseudocode describing the Muzero framework. Neural computations are implemented with Tensorflow.
You can easily train your own MuZero, more specifically for one player and non-image based environments (such as CartPole). If you wish to train Muzero on other kinds of environments, this codebase can be used with slight modifications.
DISCLAIMER: this code is early research code. What this means is:
- Silent bugs may exist.
- It may not work reliably on other environments or with other hyper-parameters.
- The code quality and documentation are quite lacking, and much of the code might still feel "in-progress".
- The training and testing pipeline is not very advanced.
We run this code using:
- Conda 4.7.12
- Python 3.7
- Tensorflow 2.0.0
- Numpy 1.17.3
This code must be run from the main function in muzero.py (don't forget to first configure your conda environment).
To train a model, please follow these steps:
- Create or modify an existing configuration of Muzero in
config.py. - Call the right configuration inside the main of
muzero.py. - Run the main function:
python muzero.py.
To train on a different environment than Cartpole-v1, please follow these additional steps:
1) Create a class that extends AbstractGame, this class should implement the behavior of your environment.
For instance, the CartPole class extends AbstractGame and works as a wrapper upon gym CartPole-v1.
You can use the CartPole class as a template for any gym environment.
2) This step is optional (only if you want to use a different kind of network architecture or value/reward transform).
Create a class that extends BaseNetwork, this class should implement the different networks (representation, value, policy, reward and dynamic) and value/reward transforms.
For instance, the CartPoleNetwork class extends BaseNetwork and implements fully connected networks.
3) This step is optional (only if you use a different value/reward transform).
You should implement the corresponding inverse value/reward transform by modifying the loss_value and loss_reward function inside training.py.
This implementation differ from the original paper in the following manners:
- We use fully connected layers instead of convolutional ones. This is due to the nature of our environment (Cartpole-v1) which as no spatial correlation in the observation vector.
- We don't scale the hidden state between 0 and 1 using min-max normalization. Instead we use a tanh function that maps any values in a range between -1 and 1.
- We do use a slightly simple invertible transform for the value prediction by removing the linear term.
- During training, samples are drawn from a uniform distribution instead of using prioritized replay.
- We also scale the loss of each head by 1/K (with K the number of unrolled steps). But, instead we consider that K is always constant (even if it is not always true).