Paper: Risi and Stanley, "Deep Neuroevolution of Recurrent and Discrete World Models" To appear in: Proceedings of the Conference on Genetic and Evolutionary Computation (GECCO 2019). New York, NY: ACM.
https://arxiv.org/abs/1906.08857
The code is partly based on the PyTorch implementation of "World Models" (https://github.com/ctallec/world-models).
Code requieres Python3 and PyTorch (https://pytorch.org). The rest of the requirements are included in the requirements file, to install them:
pip3 install -r requirements.txtThe world model is composed of three different components:
- A Variational Auto-Encoder (VAE)
- A Mixture-Density Recurrent Network (MDN-RNN)
- A linear Controller (C), which takes both the latent encoding and the hidden state of the MDN-RNN as input and outputs the agents action
In contrast to the original world model, all three components are trained end-to-end through evolution. To run training:
python main.pyTo test a specific genome:
python main.py --test best_1_1_G2.pAdditional arguments for the training script are:
- --folder : The directory to store the training results.
- --pop-size : The population size.
- --threads : The number of threads used for training or testing.
- --discrete : Switching a discrete version of the VAE on or off.
- --generations : The number of generations used for training.
- --setting : The setting determining the mutation operator. 0 = Mutate all three modules (VAE, MDN-RNN, C). 1 = Randomly mutate one of those three modules.
When running on a headless server, you will need to use xvfb-run to launch the controller training script. For instance,
xvfb-run -a -s "-screen 0 1400x900x24 +extension RANDR" -- python main.pyWhen running with a discrete VAE, the size of the latent vector is increased to 128 from the 32-dimensional version used for the standard VAE.
- Sebastian Risi
This project is licensed under the MIT License - see the LICENSE.md file for details