模型无关学习


Monte-Carlo & Temporal Difference; Q-learning





探索与利用

on-policy 和 off-policy

SARSA

Expected value SARSA

SARSA和Q-Learning对比

on-policy和off-policy对比

on-policy off-policy
Agent 可以选择动作 Agent 不能 选择动作
Most obvious setup Learning with exploration,playing without exploration
Agent always follows his own policy Learning from expert(expert is imperfect)
Learning from sessions(recorded data)
can’t learn from off-policy can learn from on-policy
SARSA Q-learning
more… Expected Value SARSA

经验回放

略