def main(MazeEnv):
env = MazeEnv()
env.render()
plt.pause(2)
n_epochs = 10000
robot_loc = []
steps = 0
rewards = 0.0
for i in range(n_epochs):
steps += 1
# next_action = np.random.randint(4,size = 1)
next_action, robot_loc = env.expert(robot_loc)
state_img, reward, done, _ = env.step(next_action)
rewards += reward
env.render()
print('Step = %d, rewards = %.1f, reward = %.1f, done = %d' %
(steps, rewards, reward, done),
end='\r')
if done:
print('\n')
if done:
steps = 0
rewards = 0.0
plt.pause(2)
env.reset()
import numpy as np
from MazeEnv import *
import copy
from collections import defaultdict
world_size = np.array([5, 5])
env = MazeEnv(world_size=world_size, \
gold_pos=np.array([[1, 2]]), \
bad_pos=np.array([[3, 4]]), \
max_ite=50)
env.reset()
DISCOUNT = 0.9
epsilon = 0.1
Q = np.zeros((world_size[0] * world_size[1], 4))
Q_old = copy.deepcopy(Q)
PI = np.zeros((world_size[0] * world_size[1]), dtype=np.int) # 0,1,2,3分别代表4个动作
#
Return = defaultdict(list)
ite = 0
while True: # episode循环
ite += 1
Q_old = copy.deepcopy(Q)
s = env.reset()
s_a_his = []
r_his = []
while True: # step循环
#a = PI[s] #随机策略
a = PI[s] if (np.random.rand() > epsilon) else env.random_action()
s_, r, d = env.step(a)
s_a_his.append(env.encode_s_a(s, a))
import numpy as np
from MazeEnv import *
world_size = np.array([5, 5])
env = MazeEnv(world_size=world_size, \
gold_pos=np.array([[1,2]]), \
bad_pos=np.array([[3,4]]), \
max_ite=100)
env.reset()
# 四个动作各1/4
ACTION_PROB = 0.25
DISCOUNT = 0.9
A1 = -1 * np.eye(world_size[0] * world_size[1])
A2 = np.zeros((world_size[0] * world_size[1], world_size[0] * world_size[1]))
b = np.zeros(world_size[0] * world_size[1])
for s in env.feasible_states:
for a in env.feasible_actions:
s_, r, d = env.step_state(s, a)
A2[s, s_] += ACTION_PROB * DISCOUNT
b[s] += ACTION_PROB * r
V = np.linalg.solve(A1 + A2, -b).reshape(world_size)
print(V)