Getting Started

An agent has to implement the act() method which takes the current state as input and returns an action:

from train import Agent

class RandomAgent(Agent):

    def act(self, state):
        return self.env.action_space.sample()

Create an environment using OpenAI Gym:

import gym

env = gym.make('CartPole-v0')

Initialize your agent using the environment:

agent = RandomAgent(env=env)

Now you can start training your agent (in this example, the agent acts randomly always and doesn’t learn anything):

scores = agent.train(episodes=100)

You can also visualize how the training progresses but it will slow down the process:

scores = agent.train(episodes=100, render=True)

Once you are done with the training, you can test it:

scores = agent.test(episodes=10)

Alternatively, visualize how it performs:

scores = agent.test(episodes=10, render=True)

To learn more about how to build an agent that learns see Agent documentation.

See examples directory to see implementations of some algorithms (DQN, A3C, PPO etc.) in TensorFlow.

Installation

Requirements:

• Python >= 3.6

Install from PyPI (recommended):

pip install train

Alternatively, install from source:

git clone https://github.com/marella/train.git
cd train
pip install -e .

To run examples and tests, install from source.

Other libraries such as Gym and TensorFlow should be installed separately.

Examples

To run examples, install TensorFlow and install dependencies:

pip install -e .[examples]

and run an example in examples directory:

cd examples
python PPO.py

Testing

To run tests, install dependencies:

pip install -e .[tests]

and run:

pytest tests