rl_lib

Motivation: I have always thought that the only way to truely test if you understand a concept is to see if you can build it. As such all these these algorithms are implemented studying the relevant papers and coded to test my understanding.

What I cannot create, I do not understand” - Richard Feynman

Algorithms

DQN

• Vanilla DQN
  • Paper: Human-level control through deep reinforcement learning
• Noisy DQN
  • Paper: Noisy Networks for Exploration
• Dualing DQN
  • Paper: Dueling Network Architectures for Deep Reinforcement Learning
• Double DQN
  • Paper: Deep Reinforcement Learning with Double Q-learning
• Prioritiesed Experience Replay DQN
  • Paper: Prioritized Experience Replay
• Rainbow DQN
  • Paper: Rainbow: Combining Improvements in Deep Reinforcement Learning

Policy Gradient

• Advantage Actor Critic (A2C) - single environment
  • Paper: Asynchronous Methods for Deep Reinforcement Learning
• Advantage Actor Critic (A2C) - multi environment
  • Paper: Asynchronous Methods for Deep Reinforcement Learning
• Deep Deterministic Policy Gradients
  • Paper: Continuous Control with Deep Reinforcement Learning
• Proximal Policy Optimisation (discrete and continuous)
  • Paper: Proximal Policy Optimization Algorithms

Tabular Solutions

These were mainly referenced from a really good lecture series by Colin Skow on youtube [link]. A large part was also found in the Deep Reinforcement Learning Udacity course.

• Bellman Equation
• Dynamic Programming
• Q learning

Results

DQN Pong

• Converged to an average of 17.56 after 1300 Episodes.
• Code and results can be found under DQN/7. Vanilla DQN Atari.ipynb

PPO discrete

• Solved in 409 episodes
• Code and results can be found under Policy Gradient/5. PPO.ipynb

DDPG Continuous

• Converged to ~ -270 after a 100 episodes
• Code and results can be found under Policy Gradient/4. DDPG.ipynb.ipynb

Todo

• General Advantage Estimator
• Pull Policy Gradient algorithms into seperate files
• Curiousity Driven Exploration
• HER (Hindsight Experience Replay)
• Recurrent networks in PPO and DDPG

Credits

Whilst I tried to code everything directly from the papers, it wasn't always easy to understand what I was doing wrong when the algorithm just wouldn't train or I got runtime errors. As such I used the following repositories as references.

• https://github.com/nikhilbarhate99/PPO-PyTorch.git
• https://github.com/hermesdt/reinforcement-learning.git