jcwleo / Reinforcement_Learning

강화학습에 대한 기본적인 알고리즘 구현

Reinforcement Learning

여러 환경에 적용해보는 강화학습 예제(파이토치로 옮기고 있습니다)

Here is my new Repo for Policy Gradient!!

---

[Breakout / Use DQN(Nature2015)]

---

1. Q-Learning / SARSA

• FrozenLake(Gridword)
  • Deterministic Q-Learning
  • Add Exploration & Discounted Factor
  • Stochastic Q-Learning
• WindyGridWorld(in Sutton's book)
  • Q-Learning / SARSA

2. Q-Network (Action-Value Function Approximation)

• FrozenLake(Gridword)
• CartPole(Classic Control)

3. DQN

DQN(NIPS2013)은 (Experience Replay Memory / CNN) 을 사용.

• CartPole(Classic Control) - Cartpole 같은 경우에는 CNN을 사용하지 않고 센서 정보를 통해서 학습

DQN(Nature2015)은 (Experience Replay Memory / Target Network / CNN) 을 사용

• CartPole(Classic Control)
• Breakout(atari)
• Breakout(atari)
  • this code is made by pytorch and more efficient memory and train

5. Vanilla Policy Gradient(REINFORCE)

• CartPole(Classic Control)
• Pong(atari)
• Breakout(atari)

6. Advantage Actor Critic

• episodic
  • CartPole(Classic Control)
  • Pong(atari)
• one-step
  • CartPole(Classic Control)
• n-step
  • CartPole(Classic Control)

7. Deep Deterministic Policy Gradient

• Pendulum(Classic Control)

8. Parallel Advantage Actor Critic(is called 'A2C' in OpenAI)

• CartPole(Classic Control)(used a single thread instead of multi thread)
• CartPole(Classic Control)(used multiprocessing in pytorch)
• Super Mario Bros(used multiprocessing in pytorch)

9. C51(Distributional RL)

• DDQN
  • CartPole(Classic Control)

10. PPO(Proximal Policy Optimization)

• CartPole(Classic Control)