def __init__(self):
self.train_episode = 1000
self.r = False # render or not
self.u = False # update or not
self.env = envR.envR(rows=10, cols=10, n_features=10)
self.max_steps = 30 # (self.env.maze.c - 2) * (self.env.maze.r - 2)
self.brain = PolicyGradient(n_actions=4,
n_features=(self.env.maze.c *
self.env.maze.r),
learning_rate=0.0001,
reward_decay=0.95,
output_graph=False,
restore=True)
# used for evaluation
self.evaluate = Evaluate(rows=10, cols=10, start_pos=(10, 1))
self.num_fail = 0
self.num_find_target = 0
self.cost, self.density = [], [] # dp is deceptive_percentage
self.opt_cost, self.opt_dp = [], [] # optimal deceptive path
self.path = []
self.reward = []
def test(RL):
env = envR(show=False)
path, cost, density, num_find_target, opt_cost = [], [], [], 0, []
evaluate = Evaluate(rows=10, cols=10)
train = False
succ = 0
print("****************************************************")
for episode in range(100):
pre_maps = env.reset()
step = 0
evaluate.set_start(start_pos=env.agent)
evaluate.set_goals(real_pos=env.maze.food_pos[0],
fake_pos=env.maze.food_pos[1])
# print("****************************************************")
# print("EPISODE ", episode)
# start_test = time.time()
for step in range(100):
action = RL.choose_action(str(pre_maps), train)
reward, done, action_ = env.step(action)
path.append(action_)
step += 1
if done:
succ += 1
cost, density, num_find_target, opt_cost = evaluation(
evaluate, cost, density, num_find_target, opt_cost, path)
path = []
break
pre_maps = env.get_maps()
print('This is ', episode, 'cost:', step, 'succ', succ)
print('average cost:', np.mean(cost), ' average density:',
np.mean(density), ' deceptive extent:', num_find_target / succ)
print('optimal cost:', np.mean(opt_cost))
print()
from envR import envR
from RL_brain import DeepQNetwork
import time
if __name__ == "__main__":
r = 1000000
index_ = '_4'
save_list = [
10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 1000, 5000, 10000, 50000,
100000, 200000, 300000, 400000, 500000, 600000, 700000, 800000, 900000,
1000000
]
train = True
env = envR(show=False)
RL = DeepQNetwork(env.n_actions,
env.n_features,
rows=env.rows,
cols=env.cols,
learning_rate=0.00001,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=200,
memory_size=5000,
e_greedy_increment=0.0005,
output_graph=False)
step = 0
succ = 0
self.sess,
tf.train.latest_checkpoint(
'/home/yiranruan/cnn_s/data/checkpoint_dir_' + name))
# def get_steps(self):
# print(self.sess.run(self.steps))
def plot_cost(self, name):
self.f = open("./cost_2_" + name + ".txt", 'a')
self.f.write(str(self.cost_his) + '\n')
self.f.close()
if __name__ == "__main__":
from envR import envR
env = envR(True)
env.reset()
RL = DeepQNetwork(env.n_actions,
env.n_features,
rows=env.rows,
cols=env.cols,
learning_rate=0.01,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=200,
memory_size=2000,
output_graph=True)
for i in range(100):
action = input('actions: ')
reward = 0
pre_maps = env.get_maps()