g
y
7
7
7
7

周博磊老师强化学习纲领笔记第二课:MDP,Policy Iteration与Value Iteration

gym环境:FrozenLake-v0:http://gym.openai.com/envs/FrozenLake-v0/

代码来自:周博磊老师的GitHub:https://github.com/cuhkrlcourse/RLexample/tree/master/MDP

环境如下:

SFFF       (S: starting point, safe)
FHFH       (F: frozen surface, safe)
FFFH       (H: hole, fall to your doom)
HFFG       (G: goal, where the frisbee is located)
  • 环境解释:冰封湖问题,智能体控制角色在网格世界中的移动。网格中的某些冰面是可行走的,而某些冰面会导致主体掉入水中。另外,智能体的移动方向是不确定的,并且仅部分取决于所选方向。(也就是如果你想向上走,你选择的动作是向上走,但是实际不一定向上走,可能会发生偏移,向左,或向右,三个方向的概率是等价的,也就是都是0.3333)代理商因找到通往目标砖的可步行路径而获得奖励。
# env.nA, 表示每个可以选择的动作的个数为4,动作空间
# env.nS, 表示状态的总数为16,状态空间
# env.P[state][a], 表示在状态state下执行动作a,返回的是prob概率, next_state下一个状态,reward奖励, done是否结束

Policy Iteration:

  • 目标:寻找一个最后策略:\(\pi\)

  • 解决方法:不断的迭代Bellman expectation backup(下面的公式5)

  • Policy Iteration algorithm:

    At each iteration t+1

    update \(v_{t+1}(s)\) from \(v_t(s')\) for all states \(s \in S\) where \(s'\) is a successor states of s

    \(v_{t+1}(s)=\sum_{a \in A}\pi(a|s)[R(s,a)+\gamma\sum_{s'\in S}P(s'|s,a)v_t(s')]\)

    Convergence: \(v_1\to v_2\to ...\to v^\pi\)

  • Iterate through the two steps:

    • Evaluate the policy \(\pi\) (computing \(v\) given current \(\pi\)),第一步:计算v函数,输入,环境,策略以及衰减因子,来计算这个策略的价值。
    • Improve the policy by acting greedily wirh respect to \(v^\pi\),第二步:改进策略policy,通过对 \(v^\pi\) (第一步通过\(\pi\)求解出来的\(v\))采取贪心的算法,来改进策略policy。
图片来自周博磊老师强化学习课程

Policy Iteration:

"""
Solving FrozenLake environment using Policy-Iteration.
Adapted by Bolei Zhou for IERG6130. Originally from Moustafa Alzantot (malzantot@ucla.edu)
"""
import numpy as np
import gym

RENDER=False
GAMMA=1.0

# 计算策略policy跑一个回合的奖励:输入环境,策略以及衰减因子,跑一个回合,返回奖励值
def run_episode(env, policy, gamma = GAMMA, render = False):
    """ Runs an episode and return the total reward """
    obs = env.reset()
    # 重置环境
    total_reward = 0
    step_idx = 0
    while True:
        if render:
            env.render()
        # 如果想看环境渲染的话,就设置输入的render为True,render默认为False
        obs, reward, done , _ = env.step(int(policy[obs]))
        total_reward += (gamma ** step_idx * reward)
        step_idx += 1
        if done:
            break
    return total_reward,step_idx

# 计算策略policy的平均奖励
def evaluate_policy(env, policy, gamma = GAMMA, n = 100):
    scores = [run_episode(env, policy, gamma, render=RENDER) for _ in range(n)]
    return np.mean(scores)

# 第一步:计算v函数,输入,环境,策略以及衰减因子,来计算这个策略的价值
def compute_policy_v(env, policy, gamma=GAMMA):
    """ Iteratively evaluate the value-function under policy.
    Alternatively, we could formulate a set of linear equations in iterms of v[s]
    and solve them to find the value function.
    """
    v = np.zeros(env.env.nS)
    eps = 1e-10
    # 将精度收敛到eps的时候,就停止更新
    while True:
        prev_v = np.copy(v)
        for s in range(env.env.nS):
            policy_a = policy[s]
            v[s] = sum([p * (r + gamma * prev_v[s_]) for p, s_, r, _ in env.env.P[s][policy_a]])
        if (np.sum((np.fabs(prev_v - v))) <= eps):
            # value converged
            break
    return v

# 第二步:改进策略policy,通过对old_policy_v采取贪心的算法,来改进策略policy
def extract_policy(v, gamma = GAMMA):
    """ Extract the policy given a value-function """
    policy = np.zeros(env.env.nS)
    for s in range(env.env.nS):
        q_sa = np.zeros(env.env.nA)
        for a in range(env.env.nA):
            q_sa[a] = sum([p * (r + gamma * v[s_]) for p, s_, r, _ in  env.env.P[s][a]])
        policy[s] = np.argmax(q_sa)

    return policy

# policy_iteration的主要算法
def policy_iteration(env, gamma = GAMMA):
    """ Policy-Iteration algorithm """
    policy = np.random.choice(env.env.nA, size=(env.env.nS))  # initialize a random policy

    max_iterations = 200000
    gamma = GAMMA
    for i in range(max_iterations):
        old_policy_v = compute_policy_v(env, policy, gamma)
        # 第一步:计算v函数,输入,环境,策略以及衰减因子,来计算这个策略的价值
        new_policy = extract_policy(old_policy_v, gamma)
        # 第二步:改进策略policy,通过对old_policy_v采取贪心的算法,来改进策略policy
        if (np.all(policy == new_policy)):
            # 如果policy已经不在发生改变了,也就是收敛了,无法提升了
            print ('Policy-Iteration converged at step %d.' %(i+1))
            break
        policy = new_policy
    return policy

if __name__ == '__main__':
    env_name  = 'FrozenLake-v0' # 'FrozenLake4x4-v0'
    env = gym.make(env_name)

    optimal_policy = policy_iteration(env, gamma = GAMMA)
    scores = evaluate_policy(env, optimal_policy, gamma = GAMMA)
    print('Average scores = ', np.mean(scores))

    print(optimal_policy)
    total,step=run_episode(env,optimal_policy,GAMMA,True)
    print("一共走了:",step)

Value Iteration:

  • 目标:寻找一个最后策略:\(\pi\)

  • 解决方法:不断的迭代Bellman optimality backup(下面的公式5)

  • Value Iteration algorithm:

    initialize \(k =1\) and \(v_0(s)=0\) for all states \(s\)

    For \(k=1\) : \(H\)

    ​ for each state \(s\)

    ​  \(q_{k+1}(s,a)=R(s,a)+\gamma\sum_{s' \in S}P(s'|s,a)v_k(s')\)

    ​  \(v_{k+1}(s)=max_aq_{k+1}(s,a)\)

    ​ \(k \leftarrow k+1\)

    To retrieve the optimal policy after the value iteration:

    \(\pi(s)=argmax_a[R(s,a)+\gamma\sum_{s' \in S}P(s'|s,a)v_{k+1}(s')]\)

Value Iteration:

"""
Solving FrozenLake environment using Value-Itertion.
Updated 17 Aug 2020
"""
import numpy as np
import gym
from gym import wrappers
from gym.envs.registration import register

def run_episode(env, policy, gamma = 1.0, render = False):
    """ Evaluates policy by using it to run an episode and finding its
    total reward.
    args:
    env: gym environment.
    policy: the policy to be used.
    gamma: discount factor.
    render: boolean to turn rendering on/off.
    returns:
    total reward: real value of the total reward recieved by agent under policy.
    """
    obs = env.reset()
    total_reward = 0
    step_idx = 0
    while True:
        if render:
            env.render()
        obs, reward, done , _ = env.step(int(policy[obs]))
        total_reward += (gamma ** step_idx * reward)
        step_idx += 1
        if done:
            break
    return total_reward


def evaluate_policy(env, policy, gamma = 1.0,  n = 100):
    """ Evaluates a policy by running it n times.
    returns:
    average total reward
    """
    scores = [
            run_episode(env, policy, gamma = gamma, render = False)
            for _ in range(n)]
    return np.mean(scores)

def extract_policy(v, gamma = 1.0):
    """ Extract the policy given a value-function """
    policy = np.zeros(env.env.nS)
    for s in range(env.env.nS):
        q_sa = np.zeros(env.action_space.n)
        for a in range(env.action_space.n):
            for next_sr in env.env.P[s][a]:
                # next_sr is a tuple of (probability, next state, reward, done)
                p, s_, r, _ = next_sr
                q_sa[a] += (p * (r + gamma * v[s_]))
        policy[s] = np.argmax(q_sa)
    return policy


def value_iteration(env, gamma = 1.0):
    """ Value-iteration algorithm """
    v = np.zeros(env.env.nS)  # initialize value-function
    max_iterations = 100000
    eps = 1e-20
    for i in range(max_iterations):
        prev_v = np.copy(v)
        q_sa=np.zeros(env.env.nA)
        for s in range(env.env.nS):
            for a in range(env.env.nA):
                q_sa[a] = sum([p*(r + gamma * prev_v[s_]) for p, s_, r, _ in env.env.P[s][a]])
            v[s] = max(q_sa)
        if (np.sum(np.fabs(prev_v - v)) <= eps):
            print ('Value-iteration converged at iteration# %d.' %(i+1))
            break
    return v


if __name__ == '__main__':

    env_name  = 'FrozenLake-v0' # 'FrozenLake4x4-v0'
    env = gym.make(env_name)
    gamma = 1.0
    optimal_v = value_iteration(env, gamma);
    policy = extract_policy(optimal_v, gamma)
    policy_score = evaluate_policy(env, policy, gamma, n=1000)
    print('Policy average score = ', policy_score)
    print(policy)

Policy Iteration和Value Iteration的区别:

①:Policy iteration主要包括两部分:policy evaluation+policy improvement,这两部分反复迭代,直到收敛。初始化一个策略policy,对策略policy进行价值评估,然后再根据价值,重新制定最优策略,反复迭代。

②:Value iteration主要包括两部分:finding optimal value function+one policy extraction,寻找一个最优的价值函数,然后根据价值函数,指定最优的策略,因为价值函数是最优的,所以策略也是最优的。

③:策略迭代的收敛速度更快一些,在状态空间较小时,最好选用策略迭代方法。当状态空间较大时,值迭代的计算量更小一些。

Problem Bellman Equation Algorithm
Prediction Bellman Expectation Equation Iterative Policy Evaluation
Control Bellman Expectation Equation Policy Iteration
Control Bellman Optimality Equation Value Iteration

Bellman expectation Equation:(当前状态跟未来状态的一个关联,\(G_t\)展开的首项是\(R_{t+1}\)

\[v^\pi(s)=E_\pi[R_{t+1}+\gamma v^\pi(s_{t+1})|s_t=s] \tag 1 \]

\[q^{\pi}(s,a)=E_\pi[R_{t+1}+\gamma q^\pi(s_{t+1},A_{t+1})|s_t=s,A_t=a] \tag 2 \]

(3)式和(4)式象征着价值函数和q函数之间的关联

\[v^\pi(s)=\sum_{a \in A}\pi(a|s)q^\pi(s,a) \tag 3 \]

\[q^\pi(s,a)=R_s^a+\gamma \sum_{s'\in S}P(s'|s,a)v^\pi(s') \tag 4 \]

(4)式带入到(3)式中得到(象征着当前状态的价值与未来状态的价值之间的一个关联)

图片来在b站周博磊老师强化学习教学视频

\[v^\pi(s)=\sum_{a \in A}\pi(a|s)[R(s,a)+\gamma \sum_{s'\in S}P(s'|s,a)v^\pi(s')] \tag 5 \]

(3)式带入到(4)式中得到(象征着当前时刻的q函数与未来时刻的q函数之间的一个关联):

图片来在b站周博磊老师强化学习教学视频

\[q^\pi(s,a)=R(s,a)+\gamma \sum_{s'\in S}P(s'|s,a)\sum_{a' \in A}\pi(a'|s')q^\pi(s',a') \tag 6 \]

公式(5)进行backup后得到Bellman expectation backup

\[v_{t+1}(s)=\sum_{a \in A}\pi(a|s)[R(s,a)+\gamma \sum_{s' \in S}P(s'|s,a)v_t(s')] \tag 7 \]

上式转换成马尔可夫奖励过程的形式:

\[v_{t+1}(s)=R^\pi(s)+\gamma P^\pi(s'|s)v_t(s') \tag 8 \]

Bellman Optimality Equation:

\[v^\pi(s)=max_{a \in A}q^\pi(s,a) \tag 9 \]

\[q^*(s,a)=R(s,a)+\gamma \sum_{s'\in S}P(s'|s,a)v^*(s') \tag {10} \]

(10)式带入到(9)式得到:

\[v^*(s)=max_aR(s,a)+\gamma \sum_{s'\in S}P(s'|s,a)v^*(s') \tag {11} \]

(9)式带入到(10)式得到:

\[q^*(s,a)=R(s,a)+\gamma \sum_{s'\in S}P(s'|s,a)max_a'q^*(s',a') \tag {12} \]

公式(4)进行backup后得到 Bellman Optimality backup

\[v(s)=max_{a\in A}[R(s,a)+\gamma \sum_{s'\in S}P(s'|s,a)v(s')] \tag {13} \]

posted @ 2021-01-21 17:01  gy77  阅读(488)  评论(0编辑  收藏  举报