Continuous control with deep reinforcement learning

Summary

• (Deterministic) Policy Gradient 사용
  • NN(non linear function approximator)로 actor-critic을 구현하는 것은 어려움(unstable and challenging)
• DQN에서 얻을 수 있는 insight
  • Experience Replay
  • Target Network사용
    • Modified using ‘soft’ target updates
  • Batch Normalization
• Actor-Critic의 Off-policy.ver