这个基于之前的工作,如果有什么疑问也可以参照以前的文章。
- 【强化学习】Q-Learning算法详解以及Python实现【80行代码】
本文主要做了两件事情
- 将上篇文章中得弱智般的
treasure on right
问题,扩张到二维。且将teasure位置随机。并添加一些节点表示坑,即到该节点得到的奖励是负数。 - 将代码结构调整了一下,让整个项目更加清晰。已经方便复用。
其实上面的 *** 作,mofan大神在github中也有提到,Q_Learning_maze
但可能每个人的具体写法不一样,但是思想基本都是一致的。
此外,我自己也做了一些小的创新or改动:
- 将越界的奖励也设置为负数,
-1
。- Q-Learning还是比较容易陷入局部最优解。如果整张图的rewards是有正有负数的情况,在经过几次负数的rewards之后,就会导致智能体偏好在边缘疯狂试探。即原地不动,因为这样他的rewards是0。而不是一般数值上的1。而这样的选择是没有意义,简直浪费时间。本来训练就慢了,还不想办法加速一下。
- 我没有选择做图形界面。觉得除了好看没什么意义。但我也一样封装了一下,这种封装的抽象化思维还是很不错的。值得学习。
把三个文件分别按名字保存。然后直接运行下面命令即可。
python treasure_maze_main.py
util.py
import time
import numpy as np
class Maze(object):
def __init__(self, shape=None, hell_num=2):
if (shape is None) or (not isinstance(shape, (tuple, list))) or (len(shape) > 2):
shape = (5, 5)
self.shape = shape
self.map = np.zeros(shape)
self.actions = {
'u': [-1, 0],
'd': [1, 0],
'l': [0, -1],
'r': [0, 1]
}
for _ in range(hell_num):
self._random_num(shape, -1)
self._random_num(shape, 1)
self.point = None
self.refresh()
def _random_num(self, shape, v):
n = shape[0] * shape[1]
while True:
rd_num = np.random.randint(0, n - 1)
y = rd_num // shape[0]
x = rd_num % shape[0]
if self.map[x][y] == 0:
self.map[x][y] = v
break
def refresh(self):
self.point = [0, 0]
def point_check(self, point):
flags = [0, 1]
for f in flags:
if (point[f] < 0) or (point[f] >= self.shape[f]):
return False
return True
def get_env_feedback(self, A):
if A not in self.actions:
raise Exception("Wrong Action")
A = self.actions[A]
point_ = [
self.point[0] + A[0],
self.point[1] + A[1]
]
if self.point_check(point_):
self.point = point_
R = self.map[self.point[0]][self.point[1]]
done = (R != 0)
else:
R, done = -1, False
return self.point, R, done
def show_matrix(self, m):
for x in m:
print(' '.join(list(map(lambda i: str(int(i)) if not isinstance(i, str) else i, x))))
def update(self, done, episode, step, r=None):
# os.system("cls")
m = self.map.tolist()
m[self.point[0]][self.point[1]] = 'x'
self.show_matrix(m)
print("==========")
if done:
print("episode: %s; step: %s; reward: %s" % (episode, step, r))
time.sleep(3)
else:
time.sleep(0.3)
RL_Brain.py
import pandas as pd
import numpy as np
class RLBrain(object):
def __init__(self, actions, lr=0.1, gamma=0.9, epsilon=0.9):
self.actions = actions
self.q_table = pd.DataFrame(
[],
columns=self.actions
)
self.lr, self.gamma, self.epsilon = lr, gamma, epsilon
def check_state(self, s):
if s not in self.q_table.index:
self.q_table = self.q_table.append(
pd.Series(
[0] * len(self.actions),
index=self.actions,
name=s
)
)
def choose_action(self, s):
self.check_state(s)
state_table = self.q_table.loc[s, :]
if (np.random.uniform() >= self.epsilon) or (state_table == 0).all():
return np.random.choice(self.actions)
else:
return np.random.choice(state_table[state_table == np.max(state_table)].index)
def learn(self, s, s_, a, r, done):
self.check_state(s_)
q_old = self.q_table.loc[s, a]
if done:
q_new = r
else:
q_new = r + self.gamma * self.q_table.loc[s_, :].max()
self.q_table.loc[s, a] += self.lr * (q_new - q_old)
treasure_maze_main.py
from RL_Brain import RLBrain
from util import Maze
if __name__ == '__main__':
ALPHA = 0.1
GAMMA = 0.9
EPSILON = 0.9
MAX_EPISODE = 15
env = Maze(shape=(3, 4))
RL = RLBrain(actions=list(env.actions.keys()))
for episode in range(MAX_EPISODE):
env.refresh()
s = env.point
step_counter = 0
done = False
env.update(done, episode, step_counter)
while not done:
a = RL.choose_action(str(s))
s_, r, done = env.get_env_feedback(a)
RL.learn(str(s), str(s_), a, r, done)
s = s_
step_counter += 1
env.update(done, episode, step_counter, r)
欢迎分享,转载请注明来源:内存溢出
评论列表(0条)