Testbed for deep reinforcement learning algorithms.
Reinforcement learning is a fundamental problem in artificial intelligence. In this setting, an agent interacts with an environment in order to maximize a reward. For example, we show our bot pixel screens of a game and want it to choose actions that result in a high score.
Mindpark is an environment for prototyping, testing, and comparing algorithms that do reinforcement learning. The library makes it easy to reuse part of behavior between algorithms, and monitor all kinds of metrics about your algorithms. It integrates well with TensorFlow, Theano, and other deep learning libraries, and with OpenAI's gym environments.
These are the algorithms that I implemented so far (feel free to contribute to this list):
|Deep Q-Network (DQN)||Mnih et al. 2015 (PDF)||Working consistently.|
|Double Deep Q-Network (DDQN)||Hasselt, Guez, Silver. 2015 (PDF)||Working consistently.|
|Asynchronous Advantage Actor-Critic (A3C)||Mnih et al. 2016 (PDF)||Partly working.|
|Reinforce||Williams 1992 (PDF)||Currently being tested.|
To get started, clone the repository and install dependencies:
git clone email@example.com:danijar/mindpark.git && cd mindpark sudo -H pip3 install .
An experiment compares between algorithms, hyper parameters, and environments.
To start an experiment, run (
-O turns on Python's optimizations):
python3 -O -m mindpark run definition/breakout.yaml
Videos and metrics are stored in a result directory, which is
~/experiment/mindpark/<timestamp>-breakout/ by default. You can plot
statistics during or after the simulation by fuzzy matching an the folder name:
python3 -m mindpark stats breakout
Let's take a look at what the previous command creates.
Experiments consist of interleaved phases of training and evaluation. For example, an algorithm might use a lower exploration rate in favor of exploitation while being evaluated. Therefore, we display the metrics in two rows:
This illustrates the metrics after a few episodes of training, as you can see on the horizontal axes. This small example is good for explanation. But if you want to take a look, here are the metrics of a longer experiment.
||During the first 80 episodes of training (the time when I ran
||The training cost of the neural network. It starts at episode 10 which is when the training starts, before that, DQN builds up its replay memory. We don't train the neural network during evaluation, so that plot is empty.|
||That's the Q-values that the dqn behavior sends to epsilon_greedy to act greedily on. You can see that they evolve over time: Action 4 seems to be quite good. But that's only for a short run, so we shouldn't conclude too much.|
||A histogram whether the current action was chosen randomly or greedy wrt the predicted Q-values. During training, epsilon is annealed, so you see a shift in the distribution. During testing, epsilon is always 0.05, so not many random actions there.|
The metric names are prefixed by the classes they come from. That's because algorithms are composed of reusable partial behaviors. See the Algorithms section for details.
Definitions are YAML files that contain all you need to run or reproduce an experiment:
epochs: 100 test_steps: 4e4 repeats: 5 envs: - Pong-v0 - Breakout-v0 algorithms: - name: LSTM-A3C (3 layers) type: A3C train_steps: 8e5 config: network: lstm_three_layers initial_learning_rate: 2e-4 - name: DQN (Mnih et al. 2015) type: DQN train_steps: 2e5 - name: Random type: Random train_steps: 0
Each algorithm will be trained on each environment for the specified number of repeats. A simulation is divided into epochs that consist of a training and an evaluation phase.
To implement your own algorithm, subclass
refer to the existing algorithms for details, and ask if you have questions.
Algorithms are composed of partial behaviors that can do preprocessing,
exploration, learning, and more. To create a reusable chunk of behavior,
There are quite a few existing behaviors that you can import from
mindpark.step and reuse in your algorithms. For more details, please look at
the according Python files or open an issue. Current behaviors include:
Mindpark is a Python 3 package, and there are no plans to support Python 2.
sudo apt-get install -y gtk2.0-dev libsdl2-dev libfluidsynth-dev libopenal-dev libboost-all-dev sudo -H python3 -c "import gym_pull; gym_pull.pull('github.com/ppaquette/gym-doom')"
TensorFlow is only needed for the existing algorithms. You are free to use your libraries of choice to implement your own algorithms.
Your pull request is very welcome. I will set up a contributors file in that case, and you can choose if and how you want to be listed.
Please follow the existing code style, and run unit tests and the integration test after changes:
python3 setup.py test python3 -m mindpark run definition/test.yaml -x
Feel free to reach out at firstname.lastname@example.org or open an issue here on Github if you have any questions.