def main():
env = gym.make('FrozenLake-v0')
rewardWindow = [0 for _ in range(100)]
qtab = qTable(env.observation_space.n, env.action_space.n)
# Excercise 5 specific (load qTable from ex3)
ex3qtab = qTable(env.observation_space.n, env.action_space.n)
f = open("ex3qtable.json", 'r').read()
ex3qtab.table = json.loads(f)
epsilon = 0.1
ep = []
rew = []
for i_episode in range(8000):
observation = env.reset()
accumulatedReward = 0
for t in range(10000):
#Render enviorment
#env.render()
#Select action
action = epsilonGreedy(epsilon, env, observation, qtab)
#Perform action
prevObs = observation
observation, reward, done, info = env.step(action)
accumulatedReward += reward
#Update Q
oldQ = qtab.getQ(prevObs, action)
currQ = qtab.getQ(observation, ex3qtab.getMaxQAction(observation)) # ex5 specific
newQ = oldQ + LEARNING_RATE*(reward + DISCOUNT*currQ - oldQ)
qtab.setQ(prevObs, action, newQ)
#Check if episode is done
if done:
rewardWindow[i_episode % 100] = accumulatedReward
ep.append(i_episode)
break
#Decrease exploration rate
epsilon *= 0.9995 # ends up at e = 0.002 after 8000 iterations
windowAvg = 0
for i in rewardWindow:
windowAvg += i
rew.append(windowAvg/100)
print(i_episode, " ", windowAvg, end='\r')
if windowAvg >= 78:
break
plt.plot(ep, rew)
plt.xlabel('episode')
plt.ylabel('reward')
plt.title('Frozen Lake on policy')
plt.grid(True)
plt.savefig("op.png")
plt.show()
def main():
env = gym.make('FrozenLake-v0')
rewardWindow = [0 for _ in range(100)]
qtab = qTable(env.observation_space.n, env.action_space.n)
epsilon = 0.1
ep = []
rew = []
for i_episode in range(8000):
observation = env.reset()
accumulatedReward = 0
for t in range(10000):
#Render enviorment
#env.render()
#Select action
action = epsilonGreedy(epsilon, env, observation, qtab)
#Perform action
prevObs = observation
observation, reward, done, info = env.step(action)
accumulatedReward += reward
#Update Q
oldQ = qtab.getQ(prevObs, action)
maxCurrQ = qtab.getMaxQ(observation)
newQ = oldQ + LEARNING_RATE * (reward + DISCOUNT * maxCurrQ - oldQ)
qtab.setQ(prevObs, action, newQ)
#Check if episode is done
if done:
rewardWindow[i_episode % 100] = accumulatedReward
ep.append(i_episode)
break
#Decrease exploration rate
epsilon *= 0.9995 # ends up at e = 0.002 after 8000 iterations
windowAvg = 0
for i in rewardWindow:
windowAvg += i
print(i_episode, " ", windowAvg, end='\r')
rew.append(windowAvg / 100)
if windowAvg >= 78:
break
plt.plot(ep, rew)
plt.xlabel('episode')
plt.ylabel('reward')
plt.title('Frozen Lake Q learning')
plt.grid(True)
plt.savefig("qlrn.png")
plt.show()
"""
Export qtable to json
"""
f = open("ex3qtable.json", 'w')
f.write(json.dumps(qtab.table))
f.close()
def main():
env = gym.make('FrozenLake-v0')
rewardWindow = [0 for _ in range(100)]
qtab = qTable(env.observation_space.n, env.action_space.n)
epsilon = 0.1
ep = []
rew = []
for i_episode in range(8000):
observation = env.reset()
action = epsilonGreedy(epsilon, env, observation, qtab)
accumulatedReward = 0
for t in range(100):
#Render enviorment
#env.render()
#Perform action
prevObs = observation
observation, reward, done, info = env.step(action)
accumulatedReward += reward
#Select action
prevAct = action
action = epsilonGreedy(epsilon, env, observation, qtab)
#Update Q
oldQ = qtab.getQ(prevObs, prevAct)
actQ = qtab.getQ(observation, action)
newQ = oldQ + LEARNING_RATE * (reward + DISCOUNT * actQ - oldQ)
qtab.setQ(prevObs, prevAct, newQ)
#Check if episode is done
if done:
rewardWindow[i_episode % 99] = accumulatedReward
ep.append(i_episode)
break
#Decrease exploration rate
epsilon *= 0.9995
windowAvg = 0
for i in rewardWindow:
windowAvg += i
rew.append(windowAvg / 100)
print(i_episode, " ", windowAvg, end='\r')
if windowAvg >= 78:
break
plt.plot(ep, rew)
plt.xlabel('episode')
plt.ylabel('reward')
plt.title('Frozen Lake SARSA')
plt.grid(True)
plt.savefig("sarsa.png")
plt.show()
def main():
env = gym.make('Taxi-v1')
rewardWindow = [0 for _ in range(100)]
qtab = qTable(env.observation_space.n, env.action_space.n)
epsilon = 0.1
ep = []
rew = []
for i_episode in range(80000):
observation = env.reset()
accumulatedReward = 0
for t in range(10000):
#Render enviorment
env.render()
#Select action
action = epsilonGreedy(epsilon, env, observation, qtab)
#Perform action
prevObs = observation
observation, reward, done, info = env.step(action)
accumulatedReward += reward
#Update Q
oldQ = qtab.getQ(prevObs, action)
maxCurrQ = qtab.getMaxQ(observation)
newQ = oldQ + LEARNING_RATE * (reward + DISCOUNT * maxCurrQ - oldQ)
qtab.setQ(prevObs, action, newQ)
#Check if episode is done
if done:
rewardWindow[i_episode % 100] = accumulatedReward
ep.append(i_episode)
rew.append(accumulatedReward)
break
#Decrease exploration rate
epsilon *= 0.9995
windowAvg = 0
for i in rewardWindow:
windowAvg += i
print(i_episode, " ", windowAvg, end='\r')
if windowAvg >= 970:
break
plt.plot(ep, rew)
plt.xlabel('episode')
plt.ylabel('reward')
plt.title('Taxi Q learning')
plt.grid(True)
plt.savefig("qlrn.png")
plt.show()