README.md

DeepRL

Implementations of some of the most well known Deep Reinforcement Learning algorithms, namely:

• Advantage Actor Critic (A2C), which make us of the Generalized Advantage Estimation procedure;
• Proximal Policy Optimization (PPO);
• Deep Q-Network;
• Coming soon

To run a specific algorithm with default options, run:

python3 src/<algorithm name>.py 

Some command-line options supported:

• --episodes, sets the number of episodes for the training;
• --max_steps or -s, sets the maximum length of a trajectory;
• --gamma or -g, sets the discount factor.

For example, to execute A2C with 500 episodes and 150-steps trajectories:

python3 src/a2c.py --episodes 500 --max_steps 150 

Advantage Actor Critic (A2C)

To test it with the OpenAI gym CartPole-v1 environment run:

python3 src/a2c.py -vv

Proximal Policy Optimization (PPO)

Specific options:

• --epsilon, to set the clip parameter;
• --epochs, number of updates of the Actor and Critic networks per episode;
• --mini_batch_size, number of samples to be used for the updates;