增强学习

1 value iteration

   for  i in max-iteration：

         for j in states：

              v[j] = max(r[j,a] + sum(p(j'|j,a)* v[j'])

2 policy iteration

   for i in max-iteration:

          policy-evaluation

         (迭代计算v [state]直至稳定,采取的action已知)

          policy-improvement

         (依次更新each state对应的action，每次取最优值)

3 model based learning

for i in max-iteration：

    1）follow policy pi, get transition list as history 

    2)  calculate reward, transition probability from history, and get P(state, prob,action,next_state)

    3) update policy using value iteration