something-original / simple

PyTorch implementation of SimPLe (Simulated Policy Learning) on the Atari 100k benchmark.

SimPLe PyTorch

PyTorch implementation of SimPLe (Simulated Policy Learning) on the Atari 100k benchmark.

Based on the paper Model-Based Reinforcement Learning for Atari.

• Installation
• How to use
• Per-environment performance

SimPLe predicting 50 frames into the future from 4 initial frames on Freeway.

Installation

This program uses python 3.7, CUDA 10.2 if enabled, and was tested on Ubuntu 20.04.1.

Run the following command to install the dependencies:

pip install torch==1.7.0 gym==0.15.7 gym[atari] opencv-python==4.4.0.42 tqdm==4.49.0 numpy==1.16.4

git clone https://github.com/openai/baselines.git
cd baselines
pip install -e .

Install wandb (optional)

You can use wandb to track your experiments:

pip install wandb

To use wandb, pass the flag --use-wandb when running the program. See How to use for more details about flags.

How to use

CUDA is enabled by default, see the following section to disable it.

To run the program, run the following command from the folder containing the simple package:

python -m simple

Disable CUDA

To disable CUDA, pass the flag --device cpu to the command line. See the next section for more information about flags.

Flags

You can pass multiple flags to the command line, a summary is printed at launch time. The most useful flags are described in the following table:

Flag	Value	Default	Description
--agents	Any positive integer	16	The number of parallel environments to train the PPO agent on
--device	Any string accepted by torch.device	cuda	Sets the PyTorch's device
--env-name	Any game name (without the suffixes) as depicted here	Freeway	Sets the gym environment

The following boolean flags are set to False if not passed to the command line:

Flag	Description
--render-evaluation	Renders the environments during evaluation
--render-training	Renders the environments during training
--use-wandb	Enables wandb to track the experiment

For example, to run the program without CUDA and to render the environments during training, run:

python -m simple --device cpu --render-training

Per-environment performance

The scores* obtained with this implementation are detailed in the following table:

Environment	Score	Paper's score	% of reported score in the original paper
Alien	558	405.2	137.7%
Freeway	22.1	20.3	108.9%
Kangaroo	640	323.1	198.1%
Kangaroo (deterministic)	466.7	481.9	96.8%
Krull	3418.2	4539.9	82.6%
MsPacman	681.3	762.8	89.3%

*Scores obtained on only one full training per environment. The scores are the maximum average cumulative reward obtained in the real environment.