Description
Introduction
In this assignment we will implement deep Q-learning, following DeepMind’s paper ([1] and [2]) that learns
to play Atari games from raw pixels. The purpose is to demonstrate the effectiveness of deep neural networks
as well as some of the techniques used in practice to stabilize training and achieve better performance. In the
process, you’ll become familiar with TensorFlow. We will train our networks on the Pong-v0 environment
from OpenAI gym, but the code can easily be applied to any other environment.
In Pong, one player scores if the ball passes by the other player. An episode is over when one of the players
reaches 21 points. Thus, the total return of an episode is between −21 (lost every point) and +21 (won
every point). Our agent plays against a decent hard-coded AI player. Average human performance is −3
(reported in [2]). In this assignment, you will train an AI agent with super-human performance, reaching at
least +10 (hopefully more!).
1
CS 234 Winter 2020: Assignment #2
0 Test Environment (6 pts)
Before running our code on Pong, it is crucial to test our code on a test environment. In this problem, you
will reason about optimality in the provided test environment by hand; later, to sanity-check your code,
you will verify that your implementation is able to achieve this optimality. You should be able to run your
models on CPU in no more than a few minutes on the following environment:
• 4 states: 0, 1, 2, 3
• 5 actions: 0, 1, 2, 3, 4. Action 0 ≤ i ≤ 3 goes to state i, while action 4 makes the agent stay in the same
state.
• Rewards: Going to state i from states 0, 1, and 3 gives a reward R(i), where R(0) = 0.1, R(1) =
−0.2, R(2) = 0, R(3) = −0.1. If we start in state 2, then the rewards defind above are multiplied by
−10. See Table 1 for the full transition and reward structure.
• One episode lasts 5 time steps (for a total of 5 actions) and always starts in state 0 (no rewards at the
initial state).
State (s) Action (a) Next State (s
0
) Reward (R)
0 0 0 0.1
0 1 1 -0.2
0 2 2 0.0
0 3 3 -0.1
0 4 0 0.1
1 0 0 0.1
1 1 1 -0.2
1 2 2 0.0
1 3 3 -0.1
1 4 1 -0.2
2 0 0 -1.0
2 1 1 2.0
2 2 2 0.0
2 3 3 1.0
2 4 2 0.0
3 0 0 0.1
3 1 1 -0.2
3 2 2 0.0
3 3 3 -0.1
3 4 3 -0.1
Table 1: Transition table for the Test Environment
An example of a trajectory (or episode) in the test environment is shown in Figure 1, and the trajectory
can be represented in terms of st, at, Rt as: s0 = 0, a0 = 1, R0 = −0.2, s1 = 1, a1 = 2, R1 = 0, s2 = 2, a2 =
4, R2 = 0, s3 = 2, a3 = 3, R3 = (−0.1) · (−10) = 1, s4 = 3, a4 = 0, R4 = 0.1, s5 = 0.
Page 2 of 8
CS 234 Winter 2020: Assignment #2
Figure 1: Example of a trajectory in the Test Environment
1. (written 6 pts) What is the maximum sum of rewards that can be achieved in a single trajectory in
the test environment, assuming γ = 1? Show first that this value is attainable in a single trajectory,
and then briefly argue why no other trajectory can achieve greater cumulative reward.
1 Q-Learning (24 pts)
Tabular setting If the state and action spaces are sufficiently small, we can simply maintain a table
containing the value of Q(s, a) – an estimate of Q∗
(s, a) – for every (s, a) pair. In this tabular setting, given
an experience sample (s, a, r, s0
), the update rule is
Q(s, a) ← Q(s, a) + α
r + γ max
a0∈A
Q (s
0
, a0
) − Q (s, a)
(1)
where α > 0 is the learning rate, γ ∈ [0, 1) the discount factor.
Approximation setting Due to the scale of Atari environments, we cannot reasonably learn and store a Q
value for each state-action tuple. We will instead represent our Q values as a function ˆq(s, a; w) where w are
parameters of the function (typically a neural network’s weights and bias parameters). In this approximation
setting, the update rule becomes
w ← w + α
r + γ max
a0∈A
qˆ(s
0
, a0
; w) − qˆ(s, a; w)
∇wqˆ(s, a; w). (2)
In other words, we aim to minimize
L(w) = E
s,a,r,s0∼D “
r + γ max
a0∈A
qˆ(s
0
, a0
; w) − qˆ(s, a; w)
2
#
(3)
Target Network DeepMind’s paper [1] [2] maintains two sets of parameters, w (to compute ˆq(s, a)) and
w− (target network, to compute ˆq(s
0
, a0
)) such that our update rule becomes
w ← w + α
r + γ max
a0∈A
qˆ
s
0
, a0
; w−
− qˆ(s, a; w)
∇wqˆ(s, a; w). (4)
and the corresponding optimization objective becomes
L
−(w) = E
s,a,r,s0∼D “
r + γ max
a0∈A
qˆ
s
0
, a0
; w−
− qˆ(s, a; w)
2
#
(5)
The target network’s parameters are updated to match the Q-network’s parameters every C training iterations, and are kept fixed between individual training updates.
Page 3 of
CS 234 Winter 2020: Assignment #2
Replay Memory As we play, we store our transitions (s, a, r, s0
) in a buffer D. Old examples are deleted
as we store new transitions. To update our parameters, we sample a minibatch from the buffer and perform
a stochastic gradient descent update.
-Greedy Exploration Strategy For exploration, we use an -greedy strategy. This means that with
probability , an action is chosen uniformly at random from A, and with probability 1 − , the greedy action
(i.e., arg maxa∈A qˆ(s, a, w)) is chosen. DeepMind’s paper [1] [2] linearly anneals from 1 to 0.1 during the
first million steps. At test time, the agent chooses a random action with probability soft = 0.05.
There are several things to be noted:
• In this assignment, we will update w every learning freq steps by using a minibatch of experiences sampled from the replay buffer.
• DeepMind’s deep Q network takes as input the state s and outputs a vector of size |A|. In the Pong
environment, we have |A| = 6 actions, so ˆq(s; w) ∈ R
6
.
• The input of the deep Q network is the concatenation 4 consecutive steps, which results in an input
after preprocessing of shape (80 × 80 × 4).
We will now examine these assumptions and implement the -greedy strategy.
1. (written 3 pts) What is one benefit of representing the Q function as ˆq(s; w) ∈ R
|A|?
2. (coding 3 pts) Implement the get action and update functions in q1 schedule.py. Test your
implementation by running python q1 schedule.py.
We will now investigate some of the theoretical considerations involved in the tuning of the hyperparameter
C which determines the frequency with which the target network weights w− are updated to match the
Q-network weights w. On one extreme, the target network could be updated every time the Q-network is
updated; it’s straightforward to check that this reduces to not using a target network at all. On the other
extreme, the target network could remain fixed throughout the entirety of training.
Furthermore, recall that stochastic gradient descent minimizes an objective of the form J(w) = Ex∼D[l(x, w)]
by making sample updates of the following form:
w ← w − α∇wl(x, w)
Stochastic gradient descent has many desirable theoretical properties; in particular, under mild assumptions,
it is known to converge to a local optimum. In the following questions we will explore the conditions under
which Q-Learning constitutes a stochastic gradient descent update.
3. (written 5 pts) Consider the first of these two extremes: standard Q-Learning without a target
network, whose weight update is given by equation (2) above. Is this weight update an instance of
stochastic gradient descent (up to a constant factor of 2) on the objective L(w) given by equation (3)?
Argue mathematically why or why not.
4. (written 5 pts) Now consider the second of these two extremes: using a target network that is never
updated (i.e. held fixed throughout training). In this case, the weight update is given by equation (4)
above, treating w− as a constant. Is this weight update an instance of stochastic gradient descent (up
to a constant factor of 2) on the objective L
−(w) given by equation (5)? Argue mathematically why
or why not.
Page 4 of 8
CS 234 Winter 2020: Assignment #2
5. (written 3 pts) An obvious downside to holding the target network fixed throughout training is that
it depends on us knowing good weights for the target network a priori; but if this was the case, we
wouldn’t need to be training a Q-network at all! In light of this, together with the discussion above
regarding the convergence of stochastic gradient descent and your answers to the previous two parts,
describe the fundamental tradeoff at play in determining a good choice of C.
6. (written, 5 pts) In supervised learning, the goal is typically to minimize a predictive model’s error on
data sampled from some distribution. If we are solving a regression problem with a one-dimensional
output, and we use mean-squared error to evaluate performance, the objective writes
L(w) = E
(x,y)∼D
[(y − f(x; w))2
]
where x is the input, y is the output to be predicted from x, D is a dataset of samples from the
(unknown) joint distribution of x and y, and f(·; w) is a predictive model parameterized by w.
This objective looks very similar to the DQN objective stated above. How are these two scenarios
different? Hint: how does this dataset D differ from the replay buffer D used above?
2 Linear Approximation (24 pts)
1. (written, 3 pts) Show that Equations (1) and (2) from problem 1 above are exactly the same when
qˆ(s, a; w) = w>δ(s, a), where w ∈ R
|S||A| and δ : S × A → R
|S||A| with
[δ(s, a)]s
0
,a0 =
1 if s
0 = s, a0 = a
0 otherwise
2. (written, 3 pts) Assuming ˆq(s, a; w) takes the form specified in the previous part, compute ∇wqˆ(s, a; w)
and write the update rule for w.
3. (coding, 15 pts) We will now implement linear approximation in TensorFlow. This question will set
up the pipeline for the remainder of the assignment. You’ll need to implement the following functions
in q2 linear.py (please read through q2 linear.py):
• add placeholders op
• get q values op
• add update target op
• add loss op
• add optimizer op
Test your code by running python q2 linear.py locally on CPU. This will run linear approximation with TensorFlow on the test environment from Problem 0. Running this implementation should
only take a minute or two.
4. (written, 3 pts) Do you reach the optimal achievable reward on the test environment? Attach the
plot scores.png from the directory results/q2 linear to your writeup.
3 Implementing DeepMind’s DQN (13 pts)
1. (coding, 10 pts) Implement the deep Q-network as described in [1] by implementing get q values op
in q3 nature.py. The rest of the code inherits from what you wrote for linear approximation. Test
your implementation locally on CPU on the test environment by running python q3 nature.py.
Running this implementation should only take a minute or two.
Page 5 of 8
CS 234 Winter 2020: Assignment #2
2. (written, 3 pts) Attach the plot of scores, scores.png, from the directory results/q3 nature
to your writeup. Compare this model with linear approximation. How do the final performances
compare? How about the training time?
4 DQN on Atari (21 pts)
Reminder: Please remember to kill your VM instances when you are done using them!!
The Atari environment from OpenAI gym returns observations (or original frames) of size (210×160×3), the
last dimension corresponds to the RGB channels filled with values between 0 and 255 (uint8). Following
DeepMind’s paper [1], we will apply some preprocessing to the observations:
• Single frame encoding: To encode a single frame, we take the maximum value for each pixel color
value over the frame being encoded and the previous frame. In other words, we return a pixel-wise
max-pooling of the last 2 observations.
• Dimensionality reduction: Convert the encoded frame to grey scale, and rescale it to (80 × 80 × 1).
(See Figure 2)
The above preprocessing is applied to the 4 most recent observations and these encoded frames are stacked
together to produce the input (of shape (80×80×4)) to the Q-function. Also, for each time we decide on an
action, we perform that action for 4 time steps. This reduces the frequency of decisions without impacting
the performance too much and enables us to play 4 times as many games while training. You can refer to
the Methods section of [1] for more details.
(a) Original input (210 × 160 × 3) with RGB colors
(b) After preprocessing in grey scale of shape (80 × 80 × 1)
Figure 2: Pong-v0 environment
1. (coding and written, 5 pts). Now we’re ready to train on the Atari Pong-v0 environment.
First, launch linear approximation on pong with python q4 train atari linear.py on Azure’s
GPU. This will train the model for 500,000 steps and should take approximately an hour. Briefly qualitatively describe how your agent’s performance changes over the course of training. Do you think that
training for a larger number of steps would likely yield further improvements in performance? Explain
your answer.
2. (coding and written, 10 pts). In this question, we’ll train the agent with DeepMind’s architecture on
the Atari Pong-v0 environment. Run python q5 train atari nature.py on Azure’s GPU.
This will train the model for 5 million steps and should take around 12 hours. Attach the plot
scores.png from the directory results/q5 train atari nature to your writeup. You should
Page 6 of 8
CS 234 Winter 2020: Assignment #2
get a score of around 13-15 after 5 million time steps. As stated previously, the DeepMind paper claims
average human performance is −3.
As the training time is roughly 12 hours, you may want to check after a few epochs that your network
is making progress. The following are some training tips:
• If you terminate your terminal session, the training will stop. In order to avoid this, you should
use screen to run your training in the background.
• The evaluation score printed on terminal should start at -21 and increase.
• The max of the q values should also be increasing.
• The standard deviation of q shouldn’t be too small. Otherwise it means that all states have
similar q values.
• You may want to use Tensorboard to track the history of the printed metrics. You can monitor
your training with Tensorboard by typing the command tensorboard –logdir=results
and then connecting to ip-of-you-machine:6006. Below are our Tensorboard graphs from
one training session:
3. (written, 3 pts) In a few sentences, compare the performance of the DeepMind DQN architecture with
the linear Q value approximator. How can you explain the gap in performance?
4. (written, 3 pts) Will the performance of DQN over time always improve monotonically? Why or why
not?
5 n-step Estimators (12 pts)
We can further understand the effects of using a bootstrapping target by adopting a statistical perspective.
As seen in class, the Monte Carlo (MC) target is an unbiased estimator1 of the true state-action value, but
it suffers from high variance. On the other hand, temporal difference (TD) targets are biased due to their
dependence on the current value estimate, but they have relatively lower variance.
There exists a spectrum of target quantities which bridge MC and TD. Consider a trajectory s1, a1, r1, s2, a2, r2, . . .
obtained by behaving according to some policy π. Given a current estimate ˆq of Qπ
, let the n-step SARSA
target (in analogy to the TD target) be defined as:
rt + γrt+1 + γ
2
rt+2 + . . . + γ
n−1
rt+n−1 + γ
n
qˆ(st+n, at+n)
1Recall that the bias of an estimator is equal to the difference between the expected value of the estimator and the quantity
which it is estimating. An estimator is unbiased if its bias is zero.
Page 7 of 8
CS 234 Winter 2020: Assignment #2
(Recall that the 1-step SARSA target is given by rt + γqˆ(st+1, at+1)).
Given that the n-step SARSA target depends on fewer sample rewards than the MC estimator, it is reasonable
to expect it to have lower variance. However, the improved bias of this target over the standard (i.e. 1-step)
SARSA target may be less obvious.
1. (written, 12 pts) Prove that for a given policy π in an infinite-horizon MDP, the n-step SARSA target
is a less-biased (in absolute value) estimator of the true state-action value function Qπ
(st, at) than is
the 1-step SARSA target. Assume that n ≥ 2 and γ < 1. Further, assume that the current value
estimate ˆq is uniformly biased across the state-action space (that is, Bias(ˆq(s, a)) = Bias(ˆq(s
0
, a0
))
for all states s, s0 ∈ S and all actions a, a0 ∈ A). You need not assume anything about the specific
functional form of ˆq.
References
[1] Volodymyr Mnih et al. “Human-level control through deep reinforcement learning”. In: Nature 518.7540
(2015), pp. 529–533.
[2] Volodymyr Mnih et al. “Playing Atari With Deep Reinforcement Learning”. In: NIPS Deep Learning
Workshop. 2013.
Page 8 of 8