跳转至

强化学习

强化学习是通过奖励函数的一种探索式学习方法。其原理如下图所示：

rl

常见的强化学习方法有 [1]：

Algorithm Description Policy Action Space State Space Operator
Monte Carlo Every visit to Monte Carlo Either Discrete Discrete Sample-means
Q-learning State–action–reward–state Off-policy Discrete Discrete Q-value
SARSA State–action–reward–state–action On-policy Discrete Discrete Q-value
Q-learning - Lambda State–action–reward–state with eligibility traces Off-policy Discrete Discrete Q-value
SARSA - Lambda State–action–reward–state–action with eligibility traces On-policy Discrete Discrete Q-value
DQN Deep Q Network Off-policy Discrete Continuous Q-value
DDPG Deep Deterministic Policy Gradient Off-policy Continuous Continuous Q-value
A3C Asynchronous Advantage Actor-Critic Algorithm On-policy Continuous Continuous Advantage
NAF Q-Learning with Normalized Advantage Functions Off-policy Continuous Continuous Advantage
TRPO Trust Region Policy Optimization On-policy Continuous Continuous Advantage
PPO Proximal Policy Optimization On-policy Continuous Continuous Advantage
TD3 Twin Delayed Deep Deterministic Policy Gradient Off-policy Continuous Continuous Q-value
SAC Soft Actor-Critic Off-policy Continuous Continuous Advantage

参考资料

  1. Reinforcement learning. Wikipedia. https://en.wikipedia.org/wiki/Reinforcement_learning