MinAtar is a testbed for AI agents which implements miniaturized versions of several Atari 2600 games. MinAtar is inspired by the Arcade Learning Environment (Bellemare et. al. 2013) but simplifies the games to make experimentation with the environments more accessible and efficient. Currently, MinAtar provides analogues to five Atari games which play out on a 10x10 grid. The environments provide a 10x10xn state representation, where each of the n channels correspond to a game-specific object, such as ball, paddle and brick in the game Breakout.
To use MinAtar, you need python3 installed, make sure pip is also up to date. To run the included DQN and AC_lambda examples, you need PyTorch. To install MinAtar, please follow the steps below:
- Clone the repo:
git clone https://github.com/kenjyoung/MinAtar.gitIf you prefer running MinAtar in a virtualenv, you can do the following before step 2:
python3 -m venv venv
source venv/bin/activate
# Upgrade Pip
pip install --upgrade pip- Install MinAtar: To install the package and its dependencies
pip install .If you have any issues with automatic dependency installation, you can instead install the necessary dependencies manually and run
pip install . --no-depsTo install additional dependencies that are used by some examples (in particular PyTorch) run
pip install ".[examples]"To verify the installation is successful, run
python examples/random_play.py -g breakoutThe program will run 1000 episodes with a random policy and report the mean and standard error in the resulting returns similar to:
Avg Return: 0.5+/-0.023194827009486406The examples/random_play.py is a simple example to demonstrate how to use the module. breakout in the previous command can be replaced by one of the five available games: asterix, breakout, freeway, seaquest and space_invaders. See the Games section below for details of each game.
To play a game as a human, run examples/human_play.py as follows:
python examples/human_play.py -g <game>Use the arrow keys to move and space bar to fire. Also, press q to quit and r to reset.
Also included in the examples directory are example implementations of DQN (dqn.py) and online actor-critic with eligibility traces (AC_lambda.py).
MinAtar now includes an OpenAI Gym plugin using the Gym plugin system. If a sufficiently recent version of OpenAI gym (pip install gym==0.21.0 works) is installed, this plugin should be automatically available after installing MinAtar as normal. A gym environment can then be constructed as follows:
import gym
env = gym.make('MinAtar/<game>')where game is one of: Asterix-v0, Breakout-v0, Freeway-v0, Seaquest-v0, SpaceInvaders-v0, Asterix-v1, Breakout-v1, Freeway-v1, Seaquest-v1, SpaceInvaders-v1. For each game, v0 specifies the version with all 6 actions available (some of which are equivalent to no-op depending on the game), while v1 specifies the version which uses the minimal action set for the game. Note that the results included in this repo and the associated paper use the full action set of 6 actions.
We provide 2 ways to visualize a MinAtar environment.
The Environment class includes a simple visualizer using matplotlib in the display_state function. To use this simply call:
env.display_state(50)where env is an instance of MinAtar.Environment. The argument is the number of milliseconds to display the state before continuing execution. To close the resulting display window call:
env.close_display()This is the simplest way to visualize the environments, unless you need to handle user input during execution in which case you could use the provided GUI class.
We also include a slightly more complex GUI to visualize the environments and optionally handle user input. This GUI is used in examples/human_play.py to play as a human and examples/agent_play.py to visualize the performance of trained agents. To use the GUI you can import it in your code with:
from minatar.gui import GUIInitialize an instance of the GUI class by providing a name for the window, and the integer number of input channels for the minatar environment to be visualized. For example:
GUI(env.game_name(), env.n_channels)where env is an instance of minatar.Environment. The recommended way to use the GUI for visualizing an environment is to include all you're agent-environment interaction code in a function that looks something like this:
def func():
gui.display_state(env.state())
#One step of agent-environment interaction here
gui.update(50, func)The first argument to gui.update is the time to hold the current frame before continuing. The second argument specifies the function to call after that time has elapsed. In the example above the call to update simply calls func again, effectively continuing the agent-environment interaction loop. Note that this is not a recursive call, as the call to func in update is made in a new thread, while the execution of the current thread continues.
To begin the execution you can use:
gui.update(0, func)
gui.run()This will enter the agent environment interaction loop and then run the GUI thread, gui.run() will block until gui.quit() is called. To handle user input you can use gui.overwrite_key_handle(on_key_event, on_release_event). The arguments are functions to be called whenever a key is pressed, and released respectively. For an example of how to do this see examples/human_play.py.
The following plots display results for DQN (Mnih et al., 2015) and actor-critic (AC) with eligibility traces. Our DQN agent uses a significantly smaller network compared to that of Mnih et al., 2015. We display results for DQN with and without experience reply. Our AC agent uses a similar architecture to DQN, but does not use experience replay. We display results for two values of the trace decay parameter, 0.8 and 0.0. Each curve is the average of 30 independent runs with different random seeds. The top plots display the sensitivity of final performance to the step-size parameter, while the bottom plots display the average return during training as a function of training frames. For further information, see the paper on MinAtar available here.
Note, the currently displayed results for Seaquest are for the version in MinAtar v1.0.10 and lower, where a bug caused the oxygen bar to flash to full one step before running out. Results for the updated version may be different.
So far we have implemented analogues to five Atari games in MinAtar as follows. For each game, we include a link to a video of a trained DQN agent playing.
The player can move freely along the 4 cardinal directions. Enemies and treasure spawn from the sides. A reward of +1 is given for picking up treasure. Termination occurs if the player makes contact with an enemy. Enemy and treasure direction are indicated by a trail channel. Difficulty is periodically increased by increasing the speed and spawn rate of enemies and treasure.
The player controls a paddle on the bottom of the screen and must bounce a ball to break 3 rows of bricks along the top of the screen. A reward of +1 is given for each brick broken by the ball. When all bricks are cleared another 3 rows are added. The ball travels only along diagonals. When the ball hits the paddle it is bounced either to the left or right depending on the side of the paddle hit. When the ball hits a wall or brick, it is reflected. Termination occurs when the ball hits the bottom of the screen. The ball's direction is indicated by a trail channel.
The player begins at the bottom of the screen and the motion is restricted to travelling up and down. Player speed is also restricted such that the player can only move every 3 frames. A reward of +1 is given when the player reaches the top of the screen, at which point the player is returned to the bottom. Cars travel horizontally on the screen and teleport to the other side when the edge is reached. When hit by a car, the player is returned to the bottom of the screen. Car direction and speed is indicated by 5 trail channels. The location of the trail gives direction while the specific channel indicates how frequently the car moves (from once every frame to once every 5 frames). Each time the player successfully reaches the top of the screen, the car speeds are randomized. Termination occurs after 2500 frames have elapsed.
The player controls a submarine consisting of two cells, front and back, to allow direction to be determined. The player can also fire bullets from the front of the submarine. Enemies consist of submarines and fish, distinguished by the fact that submarines shoot bullets and fish do not. A reward of +1 is given each time an enemy is struck by one of the player's bullets, at which point the enemy is also removed. There are also divers which the player can move onto to pick up, doing so increments a bar indicated by another channel along the bottom of the screen. The player also has a limited supply of oxygen indicated by another bar in another channel. Oxygen degrades over time and is replenished whenever the player moves to the top of the screen as long as the player has at least one rescued diver on board. The player can carry a maximum of 6 divers. When surfacing with less than 6, one diver is removed. When surfacing with 6, all divers are removed and a reward is given for each active cell in the oxygen bar. Each time the player surfaces the difficulty is increased by increasing the spawn rate and movement speed of enemies. Termination occurs when the player is hit by an enemy fish, sub or bullet; or when oxygen reaches 0; or when the player attempts to surface with no rescued divers. Enemy and diver directions are indicated by a trail channel active in their previous location to reduce partial observability.
Note: MinAtar v1.0.10 and lower have a bug in Seaquest which causes the oxygen bar to flash to full one time-step before termination occured due to oxygen running out. This could have a significant impact on agents which learn from one step transitions as a full oxygen bar could either mean full oxygen or imminent termination due to no oxygen. For this reason Seaquest results obtained prior to v1.0.11 may not be consistent with results obtained from v1.0.11 onward.
The player controls a cannon at the bottom of the screen and can shoot bullets upward at a cluster of aliens above. The aliens move across the screen until one of them hits the edge, at which point they all move down and switch directions. The current alien direction is indicated by 2 channels (one for left and one for right) one of which is active at the location of each alien. A reward of +1 is given each time an alien is shot, and that alien is also removed. The aliens will also shoot bullets back at the player. When few aliens are left, alien speed will begin to increase. When only one alien is left, it will move at one cell per frame. When a wave of aliens is fully cleared, a new one will spawn which moves at a slightly faster speed than the last. Termination occurs when an alien or bullet hits the player.
If you use MinAtar in your research please cite the following:
Young, K. Tian, T. (2019). MinAtar: An Atari-Inspired Testbed for Thorough and Reproducible Reinforcement Learning Experiments. arXiv preprint arXiv:1903.03176.
In BibTeX format:
@Article{young19minatar,
author = {{Young}, Kenny and {Tian}, Tian},
title = {MinAtar: An Atari-Inspired Testbed for Thorough and Reproducible Reinforcement Learning Experiments},
journal = {arXiv preprint arXiv:1903.03176},
year = "2019"
}
Bellemare, M. G., Naddaf, Y., Veness, J., & Bowling, M. (2013). The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47, 253–279.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., . . . others (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529.
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.