In [1]:
%matplotlib inline
from matplotlib import pyplot as plt
import gym
import numpy as np
env = gym.make('FrozenLake-v0')
In [2]:
from gym.envs.registration import register
register(
id='FrozenLakeNotSlippery-v0',
entry_point='gym.envs.toy_text:FrozenLakeEnv',
kwargs={'map_name' : '4x4', 'is_slippery': False},
max_episode_steps=100,
reward_threshold=0.78, # optimum = .8196
)
env = gym.make('FrozenLakeNotSlippery-v0')
In [3]:
Q= np.zeros((env.observation_space.n, env.action_space.n))
y = 0.95
lr = 0.1
num_episodes = 2000
rList = []
alpha=np.log(0.000001)/num_episodes
for i in range(num_episodes):
lr= np.exp(alpha*i) #adaptive learning start with exploration
s = env.reset()
rAll = 0
done=False
#The Q-Table learning algorithm
while done==False:
#randomly exploration of precisely choose best action
if np.random.rand() < lr*0.1:
a = np.random.randint(env.action_space.n)
else:
a = np.argmax(Q[s,:] )
#get new state and reward from environment
s1,reward,done,_ = env.step(a)
#giving panalty to Hole and no move
if done:
r = 1.0 if reward > 0.0 else -1.0
else:
r = -0.01
#Update Q-Table with new knowledge
Q[s,a] = Q[s,a] + lr*(r + y*np.max(Q[s1,:]) - Q[s,a])
rAll += reward
s = s1
if done == True:
break
rList.append(rAll)
print "Score over time: " + str(sum(rList[-100:])/100.0)
plt.plot(np.convolve(np.ones(100),rList,"valid"))
Out[3]:
In [4]:
s = env.reset()
d=False
n=0
while d==False:
n+=1
a = np.argmax(Q[s,:])
s,r,d,x = env.step(a)
#print("%s %s %s %s"%(s,r,d,x))
env.render()
print n
In [ ]: