Exercise TD q-learning with frozenlake from gym API
Table of Contents
Introduction
In this notebook you will use Temporal-Difference learning (TD learning), especially the Q-learning algorithm to train the agent for the Frozenlake environment provided by the gym API. The main difference between TD learning and Monte Carlo methods is that instead of learning from whole episodes one goes even one more step back and learns based on a step by step approach.
To get familiar with the Frozenlake environment it is suggested to read through the first part of the notebook exercise-monte-carlo-frozenlake-gym
. Also the usage of tie-breaking argmax and policy performance testing is explaned their in more detail.
Requirements
Knowledge
To solve this notebook you should aquire knowledge about Temporal-Difference learning, especially Q-learning, agent-environment interaction, state-action values and policy improvement.
Prerequisites
Read SUT18
chapter 6 to gain knowledge about the mentioned topics and terms.
SUT18
is the standard literature for reinforcment learning and the basis for this and following notebooks.
Python Modules
import gym
import numpy as np
import matplotlib.pyplot as plt
Frozenlake environment
env = gym.make('FrozenLake-v0')
env.reset()
env.render()
Tie-breaking argmax
def random_argmax_axis1(b):
""" a random tie-breaking argmax"""
return np.argmax(np.random.random(b.shape) * (b.T==b.max(axis=1)).T, axis=1)
##$ \epsilon $-greedy policy
def epsilon_policy(s,Q_sa,env,epsilon=0.2):
if np.random.uniform(0,1) < epsilon:
action = env.action_space.sample() #exploration
else:
action = random_argmax_axis1(Q_sa)[s] #explotation
return action
Exercise
Task
Implement the Q-learning algorithm given in chapter 6.5 in SUT18
.
def q_learning(env,nb_episode,alpha=0.5,gamma=0.95):
"""
Args:
env: given environment
nb_episode: number of episodes used for training
Kwargs:
alpha: learning rate for Q-learning
gamma: discount rate
Return:
Q_sa: learned Q-table
"""
Q_sa = np.zeros((env.observation_space.n,env.action_space.n))
#for e in range(nb_episode):
#done = False
#while not done:
#generate steps
#q_learning
return NotImplementedError
Training the agent
Q_sa = q_learning(env,100000)
print(Q_sa)
Test performance after training
Since each performance test might lead to a different value accordingly to the random nature of our frozenlake environment due to its slippery condition, it is usefull to find the mean over many tests. Reaching over 0.7 means your learning was succesful.
greedy_policy = random_argmax_axis1(Q_sa)
print(greedy_policy)
policy = lambda s: greedy_policy[s]
def test_performance(policy, nb_episodes=100):
sum_returns = 0
for i in range(nb_episodes):
state = env.reset()
done = False
while not done:
action = policy(state)
state, reward, done, info = env.step(action)
if done:
sum_returns += reward
return sum_returns/nb_episodes
mean_performance_list = []
mean_performance = 0
n = 300
for i in range(n):
performance = test_performance(policy)
mean_performance_list.append(performance)
mean_performance += performance
print(mean_performance/n)
plt.hist(mean_performance_list,bins=30)
plt.grid()
plt.xlabel("Policy performance")
plt.show()
Literature
Licenses
Notebook License (CC-BY-SA 4.0)
The following license applies to the complete notebook, including code cells. It does however not apply to any referenced external media (e.g., images).
exercise-TD-q-learning-frozenlake-gym
Oliver Fischer
is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Based on a work at https://gitlab.com/deep.TEACHING.
Code License (MIT)
The following license only applies to code cells of the notebook.
Copyright 2019 Oliver Fischer
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.