Pacman seeks reward.
Should he eat or should he run?
When in doubt, Q-learn.
In this project, you will implement value iteration and Q-learning. You will test your agents first on Gridworld (similar to the examples you've seen in class), then apply them to a simulated robot controller (Crawler) and Pacman.
As in previous projects, this project includes an autograder for you to check your solutions. This can be run on all questions with the command:
It can be run for one particular question, such as q2, by:
python autograder.py -q q2
It can be run for one particular test by commands of the form:
python autograder.py -t test_cases/q2/1-bridge-grid
The code for this project contains the following files, which are available in a zip archive:
||A value iteration agent for solving known MDPs.|
||Q-learning agents for Gridworld, Crawler and Pacman.|
||A file in which to put your answers to various questions given in the project.|
||Defines methods on general MDPs.|
||Defines the base classes
||The Gridworld implementation.|
||Classes for extracting features on (state,action) pairs. Used for the approximate q-learning agent (in qlearningAgents.py).|
||Abstract class for general reinforcement learning environments. Used
||Gridworld graphical display.|
||Plug-in for the Gridworld text interface.|
||The crawler code and test harness. You will run this but not edit it.|
||GUI for the crawler robot.|
||Parses autograder test and solution files.|
||General autograding test classes.|
||Directory containing the test cases for each question.|
||Project 3 specific autograding test classes.|
What to submit: You will fill in portions of
analysis.py during the assignment. You should submit only these files. Please don't change any others.
As always, use
turnin to submit your work. Each team should
submit only one copy of the relevant files. In addition, please send me
an email with the following information:
Evaluation: Your code will be autograded for technical correctness. Please do not change the names of any provided functions or classes within the code, or you will wreak havoc on the autograder. However, the correctness of your implementation -- not the autograder's judgements -- will be the final judge of your score. Once again, I will conduct code reviews with some of you, so we can discuss your work in detail.
To get started, run Gridworld in manual control mode, which uses the arrow keys:
python gridworld.py -m
You will see a two-exit grid layout like the one used as an example in Russell and Norvig, Chapter 17. The blue dot is the agent. Note that when you press up, the agent only actually moves north 80% of the time. Such is the life of a Gridworld agent!
You can control many aspects of the simulation. A full list of options is available by running:
python gridworld.py -h
The default agent moves randomly
python gridworld.py -g MazeGrid
You should see the random agent bounce around the grid until it happens upon an exit. Not the finest hour for an AI agent.
Note: The Gridworld MDP is such that you first must enter a pre-terminal state (the double boxes shown in the GUI) and then take the special 'exit' action before the episode actually ends (in the true terminal state called
TERMINAL_STATE, which is not shown in the GUI). If you run an episode manually, your total return may be less than you expected, due to the discount rate (
-d to change; 0.9 by default).
Look at the console output that accompanies the graphical output (or use
-t for all text).
You will be told about each transition the agent experiences (to turn this off, use
As in Pacman, positions are represented by
(x,y) Cartesian coordinates
and any arrays are indexed by
the direction of increasing
y, etc. By default,
most transitions will receive a reward of zero, though you can change this
with the living reward option (
Question 1 (6 points) Write a value iteration agent in
ValueIterationAgent, which has been partially specified for you in
valueIterationAgents.py. Your value iteration agent is an offline planner, not a reinforcement learning agent, and so the relevant training option is the number of iterations of value iteration it should run (option
-i) in its initial planning phase.
ValueIterationAgent takes an MDP on construction and runs value iteration for the specified number of iterations before the constructor returns.
Value iteration computes k-step estimates of the optimal values, Vk. In addition to running value iteration, implement the following methods for
ValueIterationAgent using Vk.
computeActionFromValues(state)computes the best action according to the value function given by
computeQValueFromValues(state, action)returns the Q-value of the (state, action) pair.
Note that values, Q-values, and policy are all displayed in the GUI: values are numbers in squares, Q-values are numbers in square quarters, and policies are arrows out from each square.
Important: Use the "batch" version of value iteration where each vector Vk is computed from a fixed vector Vk-1, not the "online" version where one single weight vector is updated in place. The difference is discussed in Sutton & Barto in chapter 4.1.
Note: A policy synthesized from values of depth k (which reflect the next k rewards) will actually reflect the next k+1 rewards (i.e., you return πk+1). Similarly, the Q-values will also reflect one more reward than the values (i.e., you return Qk+1). You may assume that 100 iterations will be enough for convergence in the questions below.
You should return the synthesized policy πk+1.
Hint: Use the
util.Counter class in
util.py, which is a dictionary with a default value of zero. Methods such as
totalCount should simplify your code. However, be careful with
argMax: the actual argmax you want may be a key not in the counter!
Note: Make sure to handle the case when a state has no available actions in an MDP (think about what this means for future rewards).
The following command loads your
ValueIterationAgent, which will compute a policy and execute it 10 times. Press a key to cycle through values, Q-values, and the simulation. You should find that the value of the start state (
V(start)) and the empirical resulting average reward are quite close.
python gridworld.py -a value -i 100 -k 10
Hint: On the default BookGrid, running value iteration for 5 iterations should give you this output:
python gridworld.py -a value -i 5
Your value iteration agent will be graded on a new grid. I will check your values, q-values, and policies after fixed numbers of iterations and at convergence (e.g. after 100 iterations).
To test your implementation, run the autograder:
python autograder.py -q q1
Question 2 (1 point)
BridgeGrid is a grid world map with the a low-reward terminal state
and a high-reward terminal state separated by a narrow "bridge", on either side of which
is a chasm of high negative reward. The agent starts near the low-reward state. With the
default discount of 0.9 and the default noise of 0.2, the optimal policy does not cross
the bridge. Change only ONE of the discount and noise parameters so that the optimal policy
causes the agent to attempt to cross the bridge. Put your answer in
analysis.py. (Noise refers to how often an
agent ends up in an unintended successor state when they perform an action.) The
default corresponds to:
python gridworld.py -a value -i 100 -g BridgeGrid --discount 0.9 --noise 0.2
To check your answer, run the autograder:
python autograder.py -q q2
Question 3 (5 points) Consider
DiscountGrid layout, shown below. This grid has two
terminal states with positive payoff (in the middle row), a close exit
with payoff +1 and a distant exit with payoff +10. The bottom row of
the grid consists of terminal states with negative payoff (shown in
red); each state in this "cliff" region has payoff -10. The starting
state is the yellow square. We distinguish between two types of
paths: (1) paths that "risk the cliff" and travel near the bottom
row of the grid; these paths are shorter but risk earning a large
negative payoff, and are represented by the red arrow in the figure
below. (2) paths that "avoid the cliff" and travel along the top
edge of the grid. These paths are longer but are less likely to
incur huge negative payoffs. These paths are represented by the
green arrow in the figure below.
Give an assignment of parameter values for discount, noise, and
livingReward which produce the following optimal policy types or
state that the policy is impossible by returning the
'NOT POSSIBLE'. The default corresponds to:
python gridworld.py -a value -i 100 -g DiscountGrid --discount 0.9 --noise 0.2 --livingReward 0.0
question3e()should each return a 3-item tuple of (discount, noise, living reward) in
Note: You can check your policies in the GUI. For example, using a correct answer to 3(a), the arrow in (0,1) should point east, the arrow in (1,1) should also point east, and the arrow in (2,1) should point north.
To check your answers, run the autograder:
python autograder.py -q q3
Note that your value iteration agent does not actually learn from experience. Rather, it ponders its MDP model to arrive at a complete policy before ever interacting with a real environment. When it does interact with the environment, it simply follows the precomputed policy (i.e., it becomes a reflex agent). This distinction may be subtle in a simulated environment like a Gridword, but it's very important in the real world, where the real MDP is not available.
Question 4 (5 points) You will now write a Q-learning agent, which does very little on construction, but instead learns by trial and error from interactions with the environment through its
update(state, action, nextState, reward) method. A stub of a Q-learner is specified in
qlearningAgents.py, and you can select it with the option
'-a q'. For this question, you must implement the
computeActionFromQValues, you should break ties
randomly for better behavior. The
random.choice() function will help. In a
particular state, actions that your agent hasn't seen before still have a Q-value,
specifically a Q-value of zero, and if all of the actions that your agent has
seen before have a negative Q-value, an unseen action may be optimal.
Important: Make sure that in your
computeActionFromQValues functions, you only access Q values by calling
getQValue . This abstraction will be useful for question 8 when you override
getQValue to use features of state-action pairs rather than state-action
With the Q-learning update in place, you can watch your q-learner learn under manual control, using the keyboard:
python gridworld.py -a q -k 5 -m
-k will control the number of episodes your agent gets
to learn. Watch how the agent learns about the state it was just in, not the one it moves
to, and "leaves learning in its wake." Hint: to help with debugging, you can turn off
noise by using the
--noise 0.0 parameter (though this obviously makes
Q-learning less interesting). If you manually steer Pacman north and then east along
the optimal path for four episodes, you should see the following Q-values:
Grading: I will run your Q-learning agent and check that it learns the same Q-values and policy as my reference implementation when each is presented with the same set of examples. To check your implementation, run the autograder:
python autograder.py -q q4
Question 5 (3 points) Complete your Q-learning agent by implementing epsilon-greedy action selection in
getAction, meaning it chooses random actions epsilon of the time, and follows its current best Q-values otherwise. When choosing a random action, include the best action as one of the options. That is, select randomly from all legal actions.
python gridworld.py -a q -k 100Your final Q-values should resemble those of your value iteration agent, especially along well-traveled paths. However, your average returns will be lower than the Q-values predict because of the random actions and the initial learning phase.
You can choose an element from a list uniformly at random by calling the
You can simulate a binary variable with probability
of success by using
util.flipCoin(p), which returns
False with probability
To check your implementation, run the autograder:
python autograder.py -q q5
With no additional code, you should now be able to run a Q-learning crawler robot:
If this doesn't work, you've probably written some code too specific to the
GridWorld problem and you should make it more general to all MDPs.
This will invoke the crawling robot from class using your Q-learner. Play around with the various learning parameters to see how they affect the agent's policies and actions. Note that the step delay is a parameter of the simulation, whereas the learning rate and epsilon are parameters of your learning algorithm, and the discount factor is a property of the environment.
Question 6 (1 point) First, train a completely random Q-learner with the default learning rate on the noiseless BridgeGrid for 50 episodes and observe whether it finds the optimal policy.
python gridworld.py -a q -k 50 -n 0 -g BridgeGrid -e 1Now try the same experiment with an epsilon of 0. Is there an epsilon and a learning rate for which it is highly likely (greater than 99%) that the optimal policy will be learned after 50 iterations?
analysis.pyshould return EITHER a 2-item tuple of
(epsilon, learning rate)OR the string
'NOT POSSIBLE'if there is none. Epsilon is controlled by
-e, learning rate by
Note: Your response should not depend on the exact tie-breaking mechanism used to choose actions. This means your answer should be correct even if, for instance, the entire bridge grid world was rotated 90 degrees.
To check your answer, run the autograder:
python autograder.py -q q6
Question 7 (1 point) Time to play some Pacman! Pacman will play games in two phases.
In the first phase, training, Pacman will begin to learn about the values of positions and actions.
Because it takes a very long time to learn accurate Q-values even for tiny grids, Pacman's training games
run in quiet mode by default, with no GUI (or console) display. Once Pacman's training is complete,
he will enter testing mode. When testing, Pacman's
self.alpha will be set to 0.0, effectively stopping Q-learning and disabling exploration, in order to allow Pacman to exploit his learned policy. Test games are shown in the GUI by default. Without any code changes you should be able to run Q-learning Pacman for very tiny grids as follows:
python pacman.py -p PacmanQAgent -x 2000 -n 2010 -l smallGridNote that
PacmanQAgentis already defined for you in terms of the
QLearningAgentyou've already written.
PacmanQAgentis only different in that it has default learning parameters that are more effective for the Pacman problem (
epsilon=0.05, alpha=0.2, gamma=0.8). You will receive full credit for this question if the command above works without exceptions and your agent wins at least 80% of the last 10 runs. The autograder here tests for 80% of the last 100 runs.
python autograder.py -q q7
Hint: If your
QLearningAgent works for
crawler.py but does not seem to be learning a good policy for Pacman on
smallGrid, it may be because your
computeActionFromQValues methods do not in some cases properly consider unseen actions. In particular, because unseen actions have by definition a Q-value of zero, if all of the actions that have been seen have negative Q-values, an unseen action may be optimal.
Note: If you want to experiment with learning parameters, you can use the option
-a, for example
-a epsilon=0.1,alpha=0.3,gamma=0.7. These values will then be accessible as
self.epsilon, self.discount and
self.alpha inside the agent. (Note the name
self.discount rather than
Note: If you want to watch 10 training games to see what's going on, use the command:
python pacman.py -p PacmanQAgent -n 10 -l smallGrid -a numTraining=10
During the 2000-game training phase with the initial command above, you will see output every 100 games with statistics about how Pacman is faring. Epsilon is positive during training, so Pacman will play poorly even after having learned a good policy: this is because he occasionally makes a random exploratory move into a ghost. As a benchmark, it should take about 1,000 games before Pacman's rewards for a 100 episode segment becomes positive, reflecting that he's started winning more than losing. By the end of training, it should remain positive and be fairly high (between 100 and 350).
Make sure you understand what is happening here: the MDP state is the exact board configuration facing Pacman, with the now complex transitions describing an entire ply of change to that state. The intermediate game configurations in which Pacman has moved but the ghosts have not replied are not MDP states, but are bundled in to the transitions.
Once Pacman is done training, he should win very reliably in test games (at least 90% of the time), since now he is exploiting his learned policy.
However, you'll find that training the same agent on the seemingly simple
mediumGrid may not work well. In the Berkeley implementation (and mine as well), Pacman's average training rewards remain negative throughout training. At test time, he plays badly, probably losing all of his test games. Training will also take a long time, despite its ineffectiveness.
Pacman fails to win on larger layouts because each board configuration is a separate state with separate Q-values. He has no way to generalize that running into a ghost is bad for all positions. Obviously, this approach will not scale.
Question 8 (3 points)
Implement an approximate Q-learning agent that learns weights for features of states, where many states might share the same features. Write your implementation in the
ApproximateQAgent class in
qlearningAgents.py, which is a subclass of
Note: Approximate Q-learning assumes the existence of a feature function f(s,a) over state and action pairs, which yields a vector f1(s,a) .. fi(s,a) .. fn(s,a) of feature values. The feature functions are provided for you in
featureExtractors.py. Feature vectors are
util.Counter (like a dictionary) objects containing the non-zero pairs of features and values; all omitted features have value zero.
The approximate Q-function takes the following form
ApproximateQAgent uses the
IdentityExtractor, which assigns a single feature to every
(state,action) pair. With this feature extractor, your approximate Q-learning agent should work identically to
PacmanQAgent. You can test this with the following command:
python pacman.py -p ApproximateQAgent -x 2000 -n 2010 -l smallGrid
ApproximateQAgent is a subclass of
QLearningAgent, and it therefore shares several methods like
getAction. Make sure that your methods in
getQValue instead of accessing Q-values directly, so that when you override
getQValue in your approximate agent, the new approximate Q-values are used to compute actions.
Once you're confident that your approximate learner works correctly with the identity features, run your approximate Q-learning agent with the custom feature extractor, which can learn to win with ease:
python pacman.py -p ApproximateQAgent -a extractor=SimpleExtractor -x 50 -n 60 -l mediumGridEven much larger layouts should be no problem for your
ApproximateQAgent. (warning: this may take a few minutes to train)
python pacman.py -p ApproximateQAgent -a extractor=SimpleExtractor -x 50 -n 60 -l mediumClassic
If you have no errors, your approximate Q-learning agent should win almost every time with these simple features, even with only 50 training games.
The student autograder will run your approximate Q-learning agent and check that it learns the same Q-values and feature weights as the reference implementation when each is presented with the same set of examples.
python autograder.py -q q8
Grading: I will run your agent on the
mediumGrid layout 10 times using a fixed random seed.
You will receive 1 point if your agent wins more than 25% of its
games, 2 points if it wins more than 50% of its games, and 3 points
if it wins more than 75% of its games. You can try your agent out under these conditions with
python pacman.py -p ApproximateQAgent -a extractor=SimpleExtractor -x 50 -n 60 -l mediumGrid -q -f
Congratulations! You have a learning Pacman agent!