vformanyuk / reinforcement-learning

Collection of reinforcement learning algorithms implementations with TensorFlow2

reinforcement-learning

Some reinforcement learning algorithms implemented with Tensorflow 2

For practise I chose OpenAI gym Lunar Lander environment. It is representative and doesn't skip frames like some other envs.

• Double DQN

• Double Dueling DQN

• Double Dueling DQN with priority based importance sampling

• Double Dueling DQN with rank based importance sampling

• Vanila Policy Gradient

• Advantage Actor-Critic with MC update

• Advantage Actor-Critic with online update, N-returns backup and entropy bonus

• Advantage Actor-Critic with online update, N-returns backup implemented in replay buffer and entropy bonus

• Proximal Policy Optimization(PPO) + Generalized Advantage Estimator(GAE)

• Deep Deterministic Policy Gradient (DDPG)

• Twin Delayed Deep Deterministic policy gradient (TD3)

• Soft Actor-Critic (SAC)

• Soft Actor-Critic with Value network (alpha term regularization taken from SAC v2)

• Soft Actor-Critic with Emphasizing Recent Experience

• APE-X DPG

  • Orchestrator
  • DPG Actor
  • DPG Learner

• APE-X with Soft Actor Critic

  • Orchestrator
  • SAC Actor
  • SAC Learner

• Curiosity based on Random Network Distillation ( with Soft Actor Critic)

• Recurrent Experience Replay in Distributed Reinforcement Lerning (R2D2) with SAC.

  • Orchestrator
  • Agent
  • Learner
  • Agent buffer Responsible for collecting trajectories and is important part of whole algorithm.

  Note 1: for this experiment the famous Lunar Lander environment was altered to produce 'stacked' states. This achived by adding liner interpolated states between 'state' and 'next_state'.

  Note 2: Because original paper says nothing about behavior near trajoctory end, the simplest approach was taken - length of last trajectoy may vary, but has length of atleast 2 records.

• Regularizing Action Policies for Smooth Control implementation based on Soft Actor-Critic

• Active dendrites networks implementation arxiv paper

  • Modified LunarLander environemnt LunarLander multitask. This implementation has two tasks: original landing task and new - lift off. Last one require lander to fly off from landing pad
  • Active Dendrits and k-Winner-Takes-All layers
  • The training script

• Quantile Regression QR-DQN

• Implicit Quantile Networks IQN

• Distributional Soft Actor-Critic