If you are in a hurry, please look directly at the four bold algorithms . They occupy different ecological niches. Please choose them according to the actual task needs. In the sub-fields of reinforcement learning (multi-agent, hierarchical reinforcement learning, reverse reinforcement learning They are the basis for developing new algorithms):
- Discrete action space recommendation: Dueling DoubleDQN (D3QN)
- Continuous action space recommendation: use TD3 if you are good at parameter adjustment , use PPO or SAC if you are not good at parameter adjustment , if the training environment Reward function is written by beginners, then use PPO
Those who have not started deep reinforcement learning, please learn the following algorithms in order:
- Getting started with deep learning/machine learning, run the MNIST dataset with multiple fully connected layers
- Getting started with deep learning/deep learning framework, use the convolutional network to run the MNIST-fashion dataset
- Getting started with classic reinforcement learning Q-learning, discrete state, discrete action
- Getting started with deep reinforcement learning DQN (Deep Q-learning), continuous state, discrete action
- Getting started with deep reinforcement learning DDPG (Deep DPG), continuous state, continuous action
- After getting started, you can start using the three algorithms of TD3, SAC and PPO
The algorithms used for getting started can only be used for getting started. Do not use them in actual projects, at least use bold algorithms. Although they are not SotA (the highest level algorithm in State of the Art) in 2021, they are good enough and not easy to use. As for being too complicated**. **Some algorithms with excellent performance but complex, I will write below. To make it easier for you to understand the subfield of reinforcement learning, I will list the outstanding research results and write a short introduction. like