edwinzhng / reinforcement-learning

Implementations of deep reinforcement learning algorithms using TensorFlow 2 and OpenAI Gym

Reinforcement Learning

Implementations of deep reinforcement learning algorithms using Tensorflow 2.0 and OpenAI Gym.

Setup

Install Anaconda for Python 3.7

Create an Anaconda environment

conda create -n rl python=3.7

Activate the Anaconda environment

conda activate rl

Install Dependencies

pip install -r requirements.txt

Running An Agent

python src/main.py --agent NETWORK_NAME --env ENV_NAME

• Available agents: DQN, A2C, PPO
• Supported environments: CartPole-v1, Acrobot-v1, MountainCar-v0
• Render the OpenAI Gym environment with --render
• Enable GPUs for training with --gpu
• Normalize environment observations with --normalization
• Set number of training episodes with --episodes

References & Code Used

• Overview of actor-critic methods and A2C
• PyTorch DQN tutorial
• OpenAI Baselines for implementation reference
• Tensorflow 2 implementations of reinforcement learning algorithms

Algorithms Implemented

• Deep Q Network (DQN) - "Playing Atari with Deep Reinforcement Learning" (Mnih, 2015) and "Human-level Control Through Deep Reinforcement Learning" (Mnih, 2015)
• Advantage Actor Critic (A2C) - "Asynchronous Methods for Deep Reinforcement Learning" (Mnih, 2016)
• Proximal Policy Optimization (PPO) - "Proximal Policy Optimization Algorithms" (Schulman, 2017)