Q-Learning 核心思想

通过 Q(s,a) = r + γ·max_a' Q(s',a') 逼近最优策略

# 伪代码
while not converged:
    s = env.reset()
    while not done:
        a = ε-greedy(Q, s)
        s', r, done = env.step(a)
        Q(s,a) = Q(s,a) + α[r + γ·max_a' Q(s',a') - Q(s,a)]
        

实验对比：Q-Learning vs DQN

关键改进：经验回放