arXiv:1602.01783v1 [cs.

LG] 4 Feb 2016

Asynchronous Methods for Deep Reinforcement Learning

Volodymyr Mnih1
Adri`
a Puigdom`
enech Badia1
1,2
Mehdi Mirza
Alex Graves1
Tim Harley1
Timothy P. Lillicrap1
David Silver1
Koray Kavukcuoglu1
1
2

[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]

Google DeepMind
Montreal Institute for Learning Algorithms (MILA), University of Montreal

Abstract
We propose a conceptually simple and lightweight framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural
network controllers. We present asynchronous variants of four standard reinforcement
learning algorithms and show that parallel actor-learners have a stabilizing effect on
training allowing all four methods to successfully train neural network controllers. The
best performing method, an asynchronous variant of actor-critic, surpasses the current state-of-the-art on the Atari domain while training for half the time on a single
multi-core CPU instead of a GPU. Furthermore, we show that asynchronous actor-critic
succeeds on a wide variety of continuous motor control problems as well as on a new
task involving finding rewards in random 3D mazes using a visual input.

Introduction

Deep neural networks provide rich representations that can enable reinforcement learning (RL) algorithms to perform effectively. However, it was previously thought that the
combination of simple online RL algorithms with deep neural networks was fundamentally
unstable. Instead, a variety of solutions have been proposed to stabilize the algorithm
[Riedmiller, 2005, Mnih et al., 2013, 2015, Van Hasselt et al., 2015, Schulman et al., 2015a].
These approaches share a common idea: the sequence of observed data encountered by an
online RL agent is non-stationary, and online RL updates are strongly correlated. By storing the agents data in an experience replay memory, the data can be batched [Riedmiller,
2005, Schulman et al., 2015a] or randomly sampled [Mnih et al., 2013, 2015, Van Hasselt
et al., 2015] from different time-steps. Aggregating over memory in this way reduces nonstationarity and decorrelates updates, but at the same time limits the methods to off-policy
reinforcement learning algorithms.
Deep RL algorithms based on experience replay have achieved unprecedented success in
challenging domains such as Atari 2600. However, experience replay has several drawbacks:
1

it uses more memory and more computation per real interaction; and it requires off-policy
learning algorithms that can update from data generated by an older policy.
In this paper we provide a very different paradigm for deep reinforcement learning.
Instead of experience replay, we asynchronously execute multiple agents in parallel, on
multiple instances of the environment. This parallelism also decorrelates the agents data
into a more stationary process, since at any given time-step the parallel agents will be
experiencing a variety of different states. This simple idea enables a much larger spectrum
of fundamental on-policy RL algorithms, such as Sarsa, n-step methods, and actor-critic
methods, as well as off-policy RL algorithms such as Q-learning, to be applied robustly and
effectively using deep neural networks.
The asynchronous reinforcement learning paradigm also offers practical benefits. Whereas
previous approaches to deep reinforcement learning rely heavily on specialized hardware
such as GPUs [Mnih et al., 2015, Van Hasselt et al., 2015, Schaul et al., 2015] or massively
distributed architectures [Nair et al., 2015], our experiments run on a single machine with a
standard multi-core CPU. When applied to a variety of Atari 2600 domains, on many games
asynchronous reinforcement learning achieves better results, in far less time than previous
GPU-based algorithms, using far less resource than massively distributed approaches. Furthermore, the best of the proposed methods, asynchronous advantage actor-critic (A3C),
was also able to master a variety of continuous motor control tasks as well as learn general
strategies for exploring 3D mazes purely from visual inputs. We believe that the success
of A3C on both 2D and 3D games, discrete and continuous action spaces, as well as its
ability to train feedforward and recurrent agents makes it the most general and successful
reinforcement learning agent to date.

Related Work

The General Reinforcement Learning Architecture (Gorila) of Nair et al. [2015] performs
asynchronous training of reinforcement learning agents in a distributed setting. In Gorila,
each process contains an actor that acts in its own copy of the environment, a separate
replay memory, and a learner that samples data from the replay memory and computes
gradients of the DQN loss [Mnih et al., 2015] with respect to the policy parameters. The
gradients are asynchronously sent to a central parameter server which updates a central
copy of the model. The updated policy parameters are sent to the actor-learners at fixed
intervals. By using 100 separate actor-learner processes and 30 parameter server instances,
for a total of 130 CPU cores, Gorila was able to significantly outperform DQN over 49 Atari
games. On many games Gorila reached the score achieved by DQN over 20 times faster
than DQN. We also note that a similar way of parallelizing DQN was proposed by Chavez
et al. [2015].
In earlier work, Li and Schuurmans [2011] applied the Map Reduce framework to parallelizing batch reinforcement learning methods with linear function approximation. Parallelism was used to speed up large matrix operations but not to parallelize the collection of
experience or stabilize learning. Grounds and Kudenko [2008] proposed a parallel version of
the Sarsa algorithm that uses multiple separate actor-learners to accelerate training. Each
actor-learner learns separately and periodically sends updates to weights that have changed
significantly to the other learners using peer-to-peer communication.
2

Tsitsiklis [1994] studied convergence properties of Q-learning in the asynchronous optimization setting. These results show that Q-learning is still guaranteed to converge when
some of the information is outdated as long as outdated information is always eventually
discarded and several other technical assumptions are satisfied. Even earlier, Bertsekas
[1982] studied the related problem of distributed dynamic programming.
Another related area of work is in evolutionary methods, which are often straightforward
to parallelize by distributing fitness evaluations over multiple machines or threads [Tomassini,
1999]. Such parallel evolutionary approaches have recently been applied to some visual reinforcement learning tasks. In one example, Koutnk et al. [2014] evolved convolutional neural
network controllers for the TORCS driving simulator by performing fitness evaluations on
8 CPU cores in parallel.

3
3.1

Background
Reinforcement Learning

We consider the standard reinforcement learning setting where an agent interacts with an
environment E over a number of discrete time steps. At each time step t, the agent receives
a state st and selects an action at from some set of possible actions A according to its policy
, where is a mapping from states st to actions at . In return, the agent receives the next
state st+1 and receives a scalar reward rt . The process continues until
P thek agent reaches a
terminal state after which the process restarts. The return Rt = k=0 rt+k is the total
accumulated return from time step t with discount factor (0, 1]. The goal of the agent
is to maximize the expected return from each state st .
The action value Q (s, a) = E [Rt |st = s, a] is the expected return for selecting action
a in state s and following policy . The optimal value function Q (s, a) = max Q (s, a)
gives the maximum action value for state s and action a achievable by any policy. Similarly,
the value of state s under policy is defined as V (s) = E [Rt |st = s] and is simply the
expected return for following policy from state s.
In value-based model-free reinforcement learning methods, the action value function is
represented using a function approximator, such as a neural network. Let Q(s, a; ) be
an approximate action-value function with parameters . The updates to can be derived
from a variety of reinforcement learning algorithms. One example of such an algorithm is Qlearning, which aims to directly approximate the optimal action value function: Q (s, a)
Q(s, a; ). In one-step Q-learning, the parameters of the action value function Q(s, a; )
are learned by iteratively minimizing a sequence of loss functions, where the ith loss function
defined as


2

Q(s0 , a0 ; i1 ) Q(s, a; i )
Li (i ) = E r + max
0
a

(1)

where s0 is the state encountered after state s.

Alternatively, Sarsa [Rummery and Niranjan, 1994, Sutton and Barto, 1998] is a widely
used on-policy algorithm where the approximate action value function Q is updated by
minimizing the following loss function during the ith iteration

2
Li (i ) = E r + Q(s0 , a0 ; i1 ) Q(s, a; i ) ,
(2)
3

where a0 is the action taken by the agent in state s0 . In tabular environments, where
Q(s, a; ) is a lookup table, both Q-learning and Sarsa are known to converge to the optimal
value function Q under certain conditions [Jaakkola et al., 1994, Tsitsiklis, 1994, Singh
et al., 2000].
We refer to the above methods as one-step Q-learning and one-step Sarsa because they
update the action value Q(s, a) toward one-step returns r + maxa0 Q(s0 , a0 ; ) and r +
Q(s0 , a0 ; ) respectively. One drawback of using one-step methods is that obtaining a
reward r only directly affects the value of the state action pair s, a that led to the reward.
The values of other state action pairs are affected only indirectly through the updated
value Q(s, a). This can make the learning process slow since many updates are required the
propagate a reward to the relevant preceding states and actions.
One way of propagating rewards faster is by using n-step returns [Watkins, 1989, Peng
and Williams, 1996]. In n-step Q-learning, Q(s, a) is updated toward the n-step return
defined as
rt + rt+1 + + n1 rt+n + max n Q(st+n+1 , a).
(3)
a

This results in a single reward r directly affecting the values of n preceding state action pairs.
This makes the process of propagating rewards to relevant state-action pairs potentially
much more efficient.
In contrast to value-based methods, policy-based model-free methods directly parameterize the policy (a|s; ) and update the parameters by performing, typically approximate, gradient ascent on E[Rt ]. One example of such a method is the REINFORCE family
of algorithms due to Williams [1992]. Standard REINFORCE updates the policy parameters in the direction log (at |st ; )Rt , which is an unbiased estimate of E[Rt ]. It is
possible to reduce the variance of this estimate while keeping it unbiased by subtracting a
learned function of the state bt (st ), known as a baseline [Williams, 1992], from the return
log (at |st ; ) (Rt bt (st )) .

(4)

A learned estimate of the value function is commonly used as the baseline bt (st ) V (st )
leading to a much lower variance estimate of the policy gradient. When an approximate
value function is used as the baseline, the quantity Rt bt used to scale the policy gradient
can be seen as an estimate of the advantage of action at in state st , or A(at , st ) = Q(at , st )
V (st ), because Rt is an estimate of Q (at , st ) and bt is an estimate of V (st ). This approach
can be viewed as an actor-critic architecture where the policy is the actor and the baseline
bt is the critic [Sutton and Barto, 1998, Degris et al., 2012].

3.2

Deep Q Networks

Temporal difference learning methods, such as Q-learning, have been known to diverge
when used with nonlinear function approximators [Tsitsiklis and Roy, 1997]. The recently
introduced variant of Q-learning for training Deep Q Networks [Mnih et al., 2015] made
use of two techniques for avoiding such divergences in practice. First, an experience replay
memory mechanism due to Lin [1993] was used to perform Q-learning updates on random
samples of past experience instead on the most recent samples of experience. Experience
replay reduces the correlations between successive updates applied to the network thereby
4

making the training data less non-stationary. Second, the network used for computing
Q-learning targets was held fixed for intervals of several thousand updates, after which
it would be updated with the current weights of Q(s, a; ). This technique of employing
a target network reduces the correlations between the target and the predicted Q-values,
again making the training problem less non-stationary. The loss function minimized by
DQN then takes the form


Q(s , a ; ) Q(s, a; )
L() = Es,a,r,s0 D r + max
0
a

2
,

(5)

where D is the experience replay memory and are the parameters of the target network.
Both experience replay and the target network were empirically shown to be important for
obtaining the best policies on a number of Atari games, but as discussed earlier, the replay
memory can have substantial memory requirements.

Asynchronous Lock-Free Reinforcement Learning

We now present multi-threaded asynchronous variants of one-step Sarsa, one-step Q-learning,

n-step Q-learning, and advantage actor-critic. The aim in designing these methods was to
find RL algorithms that can train deep neural network policies reliably and without large
resource requirements. While the underlying RL methods are quite different, with actorcritic being an on-policy policy search method and Q-learning being an off-policy valuebased method, we use two main ideas to make all four algorithms practical given our design
goal.
First, we use asynchronous actor-learners as proposed in the Gorila framework [Nair
et al., 2015], but instead of using separate machines and a parameter server, we use multiple threads on a single machine. Keeping the learners on a single machine removes the
communication costs incurred by sending gradients and parameters and enables us to use
Hogwild! [Recht et al., 2011] style updates for training the controllers.
Second, we make the observation that multiple actors-learners running in parallel are
likely to be exploring different parts of the environment. Moreover, one can explicitly use
different exploration policies in each actor-learner to maximize this diversity. By running
different exploration policies in different threads, the overall changes being made to the
parameters by multiple actor-learners applying online updates in parallel are likely to be
less correlated in time than a single agent applying online updates. Hence, we do not
use a replay memory and rely on parallel actors employing different exploration policies to
perform the stabilizing role undertaken by experience replay in the DQN training algorithm.
In addition to stabilizing learning, using multiple parallel actor-learners has multiple
practical benefits. First, we obtain a reduction in training time that is roughly linear in
the number of parallel actor-learners. Second, since we no longer rely on experience replay
for stabilizing learning we are able to use on-policy reinforcement learning methods such
as Sarsa and actor-critic to train neural networks in a stable way. We now describe our
variants of one-step Q-learning, one-step Sarsa, n-step Q-learning and advantage actorcritic, discussing design choices specific to each algorithm.

Algorithm 1 Asynchronous one-step Q-learning - pseudocode for each actor-learner thread.

// Assume global shared parameter vector .
// Assume global shared target parameter vector .
// Assume global shared counter T = 0.
Initialize thread step counter t 0
Initialize target network weights
Initialize network gradients d 0
Get initial state s
repeat
Take action a according to the -greedy policy based on Q(s, a; )
0
Receive
 new state s and reward r
r
for terminal s0
y=
0
0

r + maxa0 Q(s , a ; )
for non-terminal s0
2

Accumulate gradients wrt : d d + (yQ(s,a;))

s = s0
T T +1
tt+1
if T mod Itarget == 0 then
Update the target network
end if
if t mod IAsyncU pdate == 0 or s is terminal then
Perform asynchronous update of using d.
Clear gradients d 0.
end if
until T > Tmax

4.1

Asynchronous one-step Q-learning

Pseudocode for our variant of Q-learning, which we call Asynchronous one-step Q-learning,
is shown in Algorithm 1. Each thread interacts with its own copy of the environment
and at each step computes a gradient of the Q-learning loss. We use a shared and slowly
changing target network in computing the Q-learning loss, as was proposed in the DQN
training method. We also accumulate gradients over multiple timesteps before they are
applied, which is similar to using minibatches. This reduces the chances of multiple actors
learners overwriting each others updates in the Hogwild! setting. Accumulating updates
over several steps also provides some ability to trade off computational efficiency for data
efficiency.
Finally, we found that giving each thread a different exploration policy helps improve
robustness. Adding diversity to exploration in this manner also generally improves performance through better exploration. While there are many possible ways of making the
exploration policies differ we experiment with using -greedy exploration with  that is
periodically sampled from some distribution of values by each thread.
While Algorithm 1 gives the pseudocode for the method used in our experiments, we also
experimented with a number of variants. For example, we experimented with using thread-

specific target networks instead of using a single shared target network as in Algorithm 1.
Another choice is which network is used for selecting actions, the model network with
parameters or the target network with parameters . However, we found that these
modification led to slightly worse results on a subset of games on the Atari domain.

4.2

Asynchronous one-step Sarsa

The asynchronous one-step Sarsa algorithm is the same as asynchronous one-step Q-learning
as given in Algorithm 1 except that it uses a different target value for Q(s, a). The target
value used by one-step Sarsa is

r
for terminal s0
y=
(6)
0
0

r + Q(s , a ; )
for non-terminal s0
where a0 is the action taken in state s0 [Sutton and Barto, 1998]. We again use a target
network and updates accumulated over multiple timesteps to stabilize learning.

4.3

Asynchronous n-step Q-learning

Pseudocode for our variant of multi-step Q-learning is shown in Algorithm 2. The algorithm
is somewhat unusual because it operates in the forward view by explicitly computing n-step
returns, as opposed to the more common backward view used by techniques like eligibility
traces [Sutton and Barto, 1998]. We found that using the forward view is easier when
training neural networks with momentum-based methods and backpropagation through
time. In order to compute a single update, the algorithm first selects actions using its
exploration policy for up to tmax steps or until a terminal state is reached. This process
results in the agent receiving up to tmax rewards from the environment since its last update.
The algorithm then computes gradients for n-step Q-learning updates for each of the stateaction pairs encountered since the last update. Each n-step update uses the longest possible
n-step return resulting in a one-step update for the last state, a two-step update for the
second last state, and so on for a total of up to tmax updates. The accumulated updates
are then applied in a single gradient step.

4.4

Asynchronous advantage actor-critic

Our asynchronous variant of actor-critic is presented in Algorithm 3. The algorithm, which

we call asynchronous advantage actor-critic (A3C), maintains a policy (at |st ; ) and an
estimate of the value function V (st ; v ). Like our variant of n-step Q-learning, our variant
of actor-critic also operates in the forward view and uses the same mix of n-step returns
to update both the policy and the value-function. The policy and the value function are
updated after every tmax actions or when a terminal state is reached. The update performed
by the algorithm can be seen as 0 log (at P
|st ; 0 )A(st , at ; , v ) where A(st , at ; , v ) is an
i
k
estimate of the advantage function given by k1
i=0 rt+i + V (st+k ; v ) V (st ; v ), where
k varies from state to state and is upper-bounded by tmax .
As with the value-based methods we rely on parallel actor-learners and accumulated
updates for improving training stability. Note that while the parameters of the policy and
v of the value function are shown as being separate for generality, we always share some
7

Algorithm 2 Asynchronous n-step Q-learning - pseudocode for each actor-learner thread.

// Assume global shared parameter vector .
// Assume global shared target parameter vector .
// Assume global shared counter T = 0.
Initialize thread step counter t 1
Initialize target network parameters
Initialize thread-specific parameters 0 =
Initialize network gradients d 0
repeat
Clear gradients d 0
Synchronize thread-specific parameters 0 =
tstart = t
Get state st
repeat
Take action at according to the -greedy policy based on Q(st , a; 0 )
Receive reward rt and new state st+1
tt+1
T T +1
untilterminal st or t tstart == tmax
0
for terminal st
R=
maxa Q(st , a; )
for non-terminal st
for i {t 1, . . . , tstart } do
R ri + R
0 2
i ,ai ; ))
Accumulate gradients wrt 0 : d d + (RQ(s
0
end for
Perform asynchronous update of using d.
if T mod Itarget == 0 then

end if
until T > Tmax
of the parameters in practice. We typically use a convolutional neural network that has
one softmax output for the policy (at |st ; ) and one linear output for the value function
V (st ; v ), with all non-output layers shared.
We also found that adding the entropy of the policy to the objective function improved
exploration by discouraging premature convergence to suboptimal deterministic policies.
This technique was originally proposed by [Williams and Peng, 1991], who found that it
was particularly helpful on tasks requiring hierarchical behavior. The gradient of the full
objective function including the entropy regularization term with respect to the policy
parameters takes the form
0 log (at |st ; 0 )(Rt V (st ; v ) + 0 H((st ; 0 ))

(7)

where H is the entropy. The hyperparameter controls the strength of the entropy regularization term.
8

Algorithm 3 Asynchronous advantage actor-critic - pseudocode for each actor-learner

thread.
// Assume global shared parameter vectors and v and global shared counter T = 0
// Assume thread-specific parameter vectors 0 and v0
Initialize thread step counter t 1
repeat
Reset gradients: d 0 and dv 0.
Synchronize thread-specific parameters 0 = and v0 = v
tstart = t
Get state st
repeat
Perform at according to policy (at |st ; 0 )
Receive reward rt and new state st+1
tt+1
T T +1
untilterminal st or t tstart == tmax
0
for terminal st
R=
0
for non-terminal st // Bootstrap from last state
V (st , v )
for i {t 1, . . . , tstart } do
R ri + R
Accumulate gradients wrt 0 : d d + 0 log (ai |si ; 0 )(R V (si ; v0 ))
Accumulate gradients wrt v0 : dv dv + (R V (si ; v0 ))2 /v0
end for
Perform asynchronous update of using d and of v using dv .
until T > Tmax

4.5

Optimization

We investigated two different optimization algorithms with our asynchronous framework

stochastic gradient descent and RMSProp. Our implementations of these algorithms do not
use any locking in order to maximize throughput when using a large number of threads.
Momentum SGD: The implementation of SGD in an asynchronous setting is relatively
straightforward and well studied [Recht et al., 2011]. Let be the parameter vector that is
shared across all threads and let i be the accumulated gradients of the loss with respect
to parameters computed by thread number i. Each thread i independently applies the
standard momentum SGD update mi = mi + (1 )i followed by mi with
learning rate , momentum and without any locks. Note that in this setting, each thread
maintains its own separate gradient and momentum vector.
RMSProp: While RMSProp [Tieleman and Hinton, 2012] has been widely used in the
deep learning literature, it has not been extensively studied in the asynchronous optimization setting. The standard non-centered RMSProp update is given by
g = g + (1 )2

,

g+
9

(8)
(9)

where all operations are performed elementwise. In order to apply RMSProp in the asynchronous optimization setting one must decide whether the moving average of elementwise
squared gradients g is shared or per-thread. We experimented with two versions of the
algorithm. In one version, which we refer to as RMSProp, each thread maintains its own
g shown in Equation 8. In the other version, which we call Shared RMSProp, the vector
g is shared among threads and is updated asynchronously and without locking. We will
show that this way of sharing the statistics greatly improves the stability of the method.
Additionally, sharing statistics among threads reduces memory requirements by using one
fewer copy of the parameter vector per thread.

Experiments

We use four different platforms for assessing the properties of the proposed framework.
First, the Arcade Learning Environment [Bellemare et al., 2012] that provides a simulator
for Atari 2600 games. This is one of the most commonly used benchmark environments for
RL algorithms. We compare against state of the art results on this environment as reported
by Van Hasselt et al. [2015], Wang et al. [2015], Schaul et al. [2015], Nair et al. [2015]
and Mnih et al. [2015]. The second environment we use is the TORCS car racing simulator [Wymann et al., 2013]. TORCS is a 3D simulator where the graphics are more realistic
compared to Atari and, additionally, understanding the physics of the car is an important
component. The third environment we use to report results is the MuJoCo [Todorov, 2015]
physics simulator for evaluating agents on continuous motor control tasks with contact dynamics. The last domain, which was only used to evaluate our best-performing agent, is a
new 3D environment called Labyrinth where the agent must learn to find rewards in randomly generated mazes from a visual input. Finally, we have carried out a detailed stability
and scalability analysis of the proposed methods.

5.1

Experimental Setup

The experiments performed on a subset of Atari games (Figures 1, 6, 7 and Table 2) as

well as the TORCS experiments (Figure 2) used the following setup. Each experiment
used 16 actor-learner threads running on a single machine and no GPUs. All methods
performed updates after every 5 actions (tmax = 5 and IU pdate = 5) and shared RMSProp
was used for optimization. The three asynchronous value-based methods used a shared
target network that was updated every 40000 frames. The Atari experiments used the same
input preprocessing as Mnih et al. [2015] and an action repeat of 4. The agents used the
network architecture from Mnih et al. [2013]. The network used a convolutional layer with
16 filters of size 8 8 with stride 4, followed by a convolutional layer with with 32 filters
of size 4 4 with stride 2, followed by a fully connected layer with 256 hidden units. All
three hidden layers were followed by a rectifier nonlinearity. The value-based methods had
a single linear output unit for each action representing the action-value. The model used
by actor-critic agents had two set of outputs a softmax output with one entry per action
representing the probability of selecting the action, and a single linear output representing
the value function.
The value based methods sampled the exploration rate  from a distribution taking
10

400
Score

Score

10000
8000
6000

300
200

20

10000

10

8000

0
10

5
10 15 20
Training time (hours)

100

20

30

5
10 15 20
Training time (hours)

6000

Q*bert
DQN
1-step Q
1-step SARSA
n-step Q
A3C

DQN
1-step Q
1-step SARSA 2000
n-step Q
A3C
0
5
10 15 20
0
Training time (hours)

Space Invaders
DQN
1-step Q
1-step SARSA
n-step Q
A3C

1800
1600
1400
1200

4000

4000
2000

12000

Score

12000

500

Pong

30

Score

14000

Breakout
DQN
1-step Q
1-step SARSA
n-step Q
A3C

600

Score

Beamrider
DQN
1-step Q
1-step SARSA
n-step Q
A3C

16000

1000
800
600
400
200

5
10 15 20
Training time (hours)

5
10 15 20
Training time (hours)

Figure 1: Learning speed comparison for DQN and the new asynchronous algorithms on five
Atari 2600 games. DQN was trained on a single Nvidia K40 GPU while the asynchronous
methods were trained using 16 CPU cores. The plots are averaged over 5 runs. In the case
of DQN the runs were for different seeds with fixed hyperparameters. For asynchronous
methods we average over the best 5 models from 50 experiments with learning rates sampled
from LogU nif orm(104 , 102 ) and all other hyperparameters fixed.
three values 1 , 2 , 3 with probabilities 0.4, 0.3, 0.3. The values of 1 , 2 , 3 were annealed
from 1 to 0.1, 0.01, 0.5 respectively over the first four million frames. Advantage actor-critic
used entropy regularization with a weight = 0.01 for all Atari and TORCS experiments.
We performed a set of 50 experiments for five Atari games and every TORCS level, each
using a different random initialization and initial learning rate. The initial learning rate was
sampled from a LogU nif orm(104 , 102 ) distribution and annealed to 0 over the course
of training. We analyze the sensitivity of the methods to the choice of learning rate in
Section 5.3.2. Note that in comparisons to prior work (Tables 1 and 3) we followed standard
evaluation protocol and used fixed hyperparameters.

5.2
5.2.1

Results
Atari 2600 Games

We first present results on a subset of Atari 2600 games to demonstrate the training speed
of the new methods. Figure 1 compares the learning speed of the DQN algorithm trained
on an Nvidia K40 GPU with the asynchronous methods trained using 16 CPU cores on
five Atari 2600 games. The results show that all four asynchronous methods we presented
can successfully train neural network controllers on the Atari domain. The asynchronous
methods tend to learn faster than DQN, with significantly faster learning on some games,
while training on only 16 CPU cores. Additionally, the results suggest that n-step methods
do indeed learn faster than one-step methods. Overall, the policy-based advantage actorcritic method significantly outperforms all three value-based methods.
We then evaluated asynchronous advantage actor-critic on 57 Atari games. In order to
compare with the state of the art in Atari game playing, we largely followed the training
and evaluation protocol of Van Hasselt et al. [2015]. Specifically, we tuned hyperparameters
11

Method
DQN (from [Nair et al., 2015])
Gorila [Nair et al., 2015]
Double DQN [Van Hasselt et al., 2015]
Dueling Double DQN [Wang et al., 2015]
Prioritized DQN [Schaul et al., 2015]
A3C, FF
A3C, FF
A3C, LSTM

Training Time
8 days on GPU
4 days, 100 machines
8 days on GPU
8 days on GPU
8 days on GPU
1 day on CPU
4 days on CPU
4 days on CPU

Mean
121.9%
215.2%
332.9%
343.8%
463.6%
344.1%
496.8%
623.0%

Median
47.5%
71.3%
110.9%
117.1%
127.6%
68.2%
116.6%
112.6%

Table 1: Mean and median human-normalized scores on 57 Atari games using the human
starts evaluation metric. Table 3 shows the raw scores for all games.

(learning rate and amount of gradient norm clipping) using a search on six Atari games
(Beamrider, Breakout, Pong, Q*bert, Seaquest and Space Invaders) and used the best
hyperparameters for all 57 games. We trained both a feedforward agent with the same
architecture as [Mnih et al., 2015, Nair et al., 2015, Van Hasselt et al., 2015] as well as a
recurrent agent with an additional 256 LSTM [Hochreiter and Schmidhuber, 1997] cells after
the final hidden layer. We additionally used the final network weights for evaluation to make
the results more comparable to the original results from Bellemare et al. [2012]. We trained
our agents for four days using 16 CPU cores, while the other agents were trained for 8 to
10 days on Nvidia K40 GPUs. Table 1 shows the average and median human-normalized
scores obtained by our agents trained by asynchronous advantage actor-critic (A3C) as well
as the current state-of-the art while Table 3 shows the scores on all games. A3C significantly
improves on state-of-the-art the average score over 57 games in half the training time of
the other methods while using only 16 CPU cores and no GPU. Furthermore, after just
one day of training, A3C matches the average human normalized score of Dueling Double
DQN as well as the median human normalized score of DQN. We note that many of the
improvements that are presented in Double DQN [Van Hasselt et al., 2015] and Dueling
Double DQN [Wang et al., 2015] can be incorporated to 1-step Q and n-step Q methods
presented in this work with similar potential improvements.
5.2.2

TORCS Car Racing Simulator

We also compared the four asynchronous methods on the TORCS 3D car racing game [Wymann
et al., 2013]. TORCS not only has more realistic graphics than Atari 2600 games, but also
requires the agent to learn the dynamics of the car it is controlling. At each step, an
agent received only a visual input in the form of an RGB image of the current frame as
well as a reward proportional to the agents velocity along the center of the track at the
agents current position. This reward structure differs considerably from most Atari games,
where the rewards are usually very sparse. We used the same neural network architecture
as the one used in the Atari experiments specified in Section 5.1. We performed experiments using four different settings the agent controlling a slow car with and without
opponent bots, and the agent controlling a fast car with and without opponent bots. The

12

Slow car, no bots

4000

4000

3000

3000

2000
1000
0
1000

10

0
1000

5000

5000

4000

4000

3000

3000

2000
1000
0
1000

10

10

Async 1-step Q
Async SARSA
Async n-step Q
Async actor-critic
Human tester
20
30
40
Training time (hours)

Async 1-step Q
Async SARSA
Async n-step Q
Async actor-critic
Human tester
20
30
40
Training time (hours)
Fast car, bots

6000

Score

Score

2000
1000

Async 1-step Q
Async SARSA
Async n-step Q
Async actor-critic
Human tester
20
30
40
Training time (hours)
Fast car, no bots

6000

Slow car, bots

5000

Score

Score

5000

2000
1000
0
1000

10

Async 1-step Q
Async SARSA
Async n-step Q
Async actor-critic
Human tester
20
30
40
Training time (hours)

Figure 2: Comparison of algorithms on the TORCS car racing simulator. Four different configurations of car speed and opponent presence or absence are shown. In each
plot, all four algorithms (one-step Q, one-step Sarsa, n-step Q and Advantage ActorCritic) are compared on score vs training time in wall clock hours. Multi-step algorithms
achieve better policies much faster than one-step algorithms on all four levels. The curves
show averages over the 5 best runs from 50 experiments with learning rates sampled from
LogU nif orm(104 , 102 ) and all other hyperparameters fixed.
results for the different game configurations comparing all four algorithms are shown in
Figure 2. Multi-step algorithms learn much faster and reach better policies on all four configurations. Moreover, the best method, Async Advantage Actor-Critic approached its best
performance after roughly 12 hours of training. Its performance reached between roughly
75% and 90% of the score obtained by a human tester on all four game configurations. A
video showing the learned driving behavior of the best performing agent can be found at
https://youtu.be/0xo1Ldx3L5Q.
5.2.3

Continuous Action Control Using the MuJoCo Physics Simulator

Finally, we also examined a set of tasks where the action space is continuous. In particular,
we look at a set of rigid body physics domains with contact dynamics where the tasks
13

Figure 3: Performance for the Mujoco continuous action domains. Scatter plot of the best
score obtained against learning rates sampled from LogU nif orm(105 , 101 ). For nearly
all of the tasks there is a wide range of learning rates that lead to good performance on the
task.
include many examples of manipulation and locomotion. These tasks were simulated in the
Mujoco physics engine. The action space for the Atari domains is naturally discrete and for
TORCS a small discretization of the action space is straightforward and was found to be
successful. However, there are many problems for which discretization of the action space
is unlikely to be a good strategy. If, for example, a problem requires controlling a system
with 10 independently controlled joint torques, then even very coarse discretization of the
action space into 5 values for each joint leads to 510 discrete actions. Because of this fact,
the DQN algorithm (or any algorithm that relies on a max operator over actions) cannot

14

Figure 4: Score per episode vs wall-clock time plots for the Mujoco domains. Each plot
shows error bars for the top 5 experiments.
easily be applied to continuous control problems with even moderately sized action spaces.
However, one of the algorithms examined here, the asynchronous advantage actor-critic,
is straightforward to apply in continuous action spaces. Since this algorithm does not rely
on the max operator over actions, all that is required to apply it to the Mujoco domains is
to ensure that the actor network outputs a vector sampled from a continuous distribution
in the appropriately sized space. Thus, in the context of the continuous action control
problems we examined only the asynchronous advantage actor-critic algorithm. Since most
of the design choices for the algorithm were made with discrete control problems in mind,
these results serve as a proof-of-concept application and could likely be improved by further
adapting the method to continuous control tasks.
15

To apply the asynchronous advantage actor-critic algorithm to the Mujoco tasks the
necessary setup is nearly identical to that used in the discrete action domains, so here we
enumerate only the differences required for the continuous action domains. The essential
elements for many of the tasks (i.e. the physics models and task objectives) are near identical
to the tasks examined in [Lillicrap et al., 2015]. However, the rewards and thus performance
are not comparable for most of the tasks due to changes made by the developers of Mujoco
which altered the contact model.
For all the domains we attempted to learn the task using the physical state as input. The
physical state consisted of the joint positions and velocities as well as the target position if
the task required a target. In addition, for three of the tasks (pendulum, pointmass2D, and
gripper) we also examined training directly from RGB pixel inputs. In the low dimensional
physical state case, the inputs are mapped to a hidden state using one hidden layer with 200
ReLU units. In the cases where we used pixels, the input was passed through two layers
of spatial convolutions without any non-linearity or pooling. In either case, the output
of the encoder layers were fed to a single layer of 128 LSTM cells. The most important
difference in the architecture is in the the output layer of the policy network. Unlike the
discrete action domain where the action output is a Softmax, here the two outputs of the
policy network are two real number vectors which we treat as the mean vector and scalar
variance 2 of a multidimensional normal distribution with a spherical covariance. To act,
the input is passed through the model to the output layer where we sample from the normal
distribution determined by and 2 . In practice, is modeled by a linear layer and 2
by a SoftPlus operation, log(1 + exp(x)), as the activation computed as a function of the
output of a linear layer. In our experiments with continuous control problems the networks
for policy network and value network do not share any parameters, though this detail is
unlikely to be crucial. Finally, since the episodes were typically at most several hundred
time steps long, we did not use any bootstrapping in the policy or value function updates
and batched each episode into a single update.
As in the discrete action case, we included an entropy cost which encouraged exploration. In the continuous case the we used a cost on the differential entropy of the normal
distribution defined by the output of the actor network, 21 (log(2 2 ) + 1), we used a constant multiplier of 104 for this cost across all of the tasks examined. The asynchronous
advantage actor-critic algorithm finds solutions for all the domains. Figure 4 shows learning
curves against wall-clock time, and demonstrates that most of the domains from states can
be solved within a few hours. All of the experiments, including those done from pixel based
observations, were run on CPU. Even in the case of solving the domains directly from pixel
inputs we found that it was possible to reliably discover solutions within 24 hours. Figure
3 shows scatter plots of the top scores against the sampled learning rates. In most of the
domains there is large range of learning rates that consistently achieve good performance
on the task.
Some of the successful policies learned by our agent can be seen in the following video
https://youtu.be/Ajjc08-iPx8.

16

Top 5 Labyrinth Agents

60

A3C

Average score

50
40
30
20
10
0

50

100

150

Training steps in millions

200

Figure 5: Training curves for the best 5 Labyrinth agents selected from a search over 50
random learning rates and entropy penalties. Training took approximately 3 days.

5.2.4

Labyrinth

We performed an additional set of experiments with A3C on a new 3D environment called

Labyrinth. The specific task we considered involved the agent learning to find rewards in
randomly generated mazes. At the beginning of each episode the agent was placed in a
new randomly generated maze consisting of rooms and corridors. Each maze contained two
types of objects that the agent was rewarded for finding apples and portals. Picking up
the agent lead to a reward of 1. Entering a portal lead to a reward of 10 after which the
agent was respawned in a new random location in the maze and all previously collected
apples were regenerated. An episode terminated after 60 seconds after which a new episode
would begin. The aim of the agent is to collect as many points as possible in the time
limit and the optimal strategy involves first finding the portal and then repeatedly going
back to it after each respawn. This task is much more challenging than the TORCS driving
domain because the agent is faced with a new maze in each episode and must learn a general
strategy for exploring random mazes.
We trained an A3C LSTM agent on this task using only 84 84 RGB images as input.
Figure 5 shows an averaged training curve for the best 5 agents we trained. The final average
score of around 50 indicates that the agent learned a reasonable strategy for exploring
random 3D maxes using only a visual input. A video showing one of the agents exploring
previously unseen mazes is included at https://youtu.be/nMR5mjCFZCw.

5.3
5.3.1

Analysis
Scalability and Data Efficiency

We now analyze the effectiveness of our proposed framework by looking at how the training
time and data efficiency changes with the number of parallel actor-learners. When using
multiple workers in parallel and updating a shared model, one would expect that in an ideal
case, for a given task and algorithm, the total number of training steps to achieve a certain
17

Method
1-step Q
1-step SARSA
n-step Q
A3C

1
1.0
1.0
1.0
1.0

Number of
2
4
3.0 6.3
2.8 5.9
2.7 5.9
2.1 3.7

threads
8
16
13.3 24.1
13.1 22.1
10.7 17.2
6.9
12.5

Table 2: The average training speedup for each method and number of threads averaged
over seven Atari games. To compute the training speed-up on a single game we measured
the time to required reach a fixed reference score using each method and number of threads.
The speedup from using n threads on a game was defined as the time required to reach
a fixed reference score using one thread divided the time required to reach the reference
score using n threads. The table shows the speedups averaged over seven Atari games
(Beamrider, Breakout, Enduro, Pong, Q*bert, Seaquest, and Space Invaders).

score would remain the same with varying numbers of workers. Therefore, the advantage
would be solely due to the ability of the system to consume more data in the same amount
of wall clock time and possibly improved exploration. Table 2 shows the training speed-up
achieved by using increasing numbers of parallel actor-learners averaged over seven Atari
games. These results show that all four methods achieve substantial speedups from using
multiple worker threads, with 16 threads leading to at least an order of magnitude speedup.
This confirms that our proposed framework scales well with the number of parallel workers,
making efficient use of resources.
Somewhat surprisingly, asynchronous one-step Q-learning and Sarsa algorithms exhibit
superlinear speedups that cannot be explained by purely computational gains. These effects
are shown more clearly in Figure 6, which shows plots of the average score against the total
number of training frames for different numbers of actor-learners and training methods on
five Atari games, and Figure 7, which shows plots of the average score against wall-clock
time. Figure 6 shows that one-step methods (one-step Q and one-step Sarsa) often require
less data to achieve a particular score when using more parallel actor-learners. While a
similar effect exists for n-step Q-learning it is less dramatic. When these gains in data
efficiency are combined with a sublinear computational speedup, n-step Q-learning achieves
a linear speedup from using multiple actor-learners while one-step Q-learning and Sarsa
achieve superlinear gains shown in Table 2. The data efficiency of asynchronous advantage
actor-critic seems to be largely unaffected by the number of parallel actor-learners. Nevertheless, asynchronous actor-critic still exhibits a substantial speedup, training over 12 times
faster using 16 actor-learners.
5.3.2

Robustness and Stability

We first analyze three asynchronous optimization algorithms by inspecting their sensitivity

to different learning rates and random network initializations. Stochastic gradient descent
is still widely used for training neural networks due to its simplicity and computational
efficiency. Although there are many extensions such as ADAGRAD [Duchi et al., 2011],

18

40

350

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

10000

250
Score

4000

10

20
30
Training epochs
Beamrider

150

Score

4000

10

20
30
Training epochs
Breakout

16000

10

20
30
Training epochs
Beamrider

12000

150

10

20
30
Training epochs
Breakout

Score

300
200

2000

100

10

20
30
Training epochs

40

10

20
30
Training epochs
Pong

10

20
30
Training epochs
Pong

20
30
Training epochs

40

Score

700

300

500

200
0

10

20
30
Training epochs
Q*bert

100

40

10

8000

10

20
30
Training epochs

40

500
400
300

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

10

20
30
Training epochs
Q*bert

100

40

1400

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

10

20
30
Training epochs
Space Invaders

40

20
30
Training epochs

40

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

1200
1000
800
600
400

2000

200
0

10

20
30
Training epochs

40

10

Figure 6: Data efficiency comparison of different numbers of actor-learners for all four
asynchronous methods on five Atari games. The x-axis shows the total number of training
epochs where an epoch corresponds to four million frames (across all threads). The y-axis
shows the average score. Each curve shows the average of the three best performing agents
from a search over 50 random learning rates. Single step methods show increased data
efficiency with increased numbers of parallel workers.

19

40

700

6000

20
30
Training epochs
Space Invaders

600

4000
A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

10

200

12000
10000

800

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

3000

20

40

500

1000

20
30
Training epochs
Space Invaders

400

1000

40

10

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

800

2000

20
10

900

4000

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

600

6000

10

100

40

2000

40

30

20
30
Training epochs
Q*bert

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

5000

30

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

400

4000

25

40

500

6000

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

20
0

10

1500

15

600

8000

2500

10

700

10000

800

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

14000

3000

10

200

40

3500

15

50
0

10

20

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

4500
4000

25

40

100

2000

40

15

15

250

6000

20
30
Training epochs
Pong

20

300

8000

10

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

200

500

10

350

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

10000

Score

200

40

20

50

12000

Score

25

40

100

2000

20
30
Training epochs
Breakout

Score

Score

6000

10

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

300

8000

500
400
300

1000

Score

20
30
Training epochs
Beamrider

2000
1500

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

20

Score

12000

10

700
600

Score

50
0

3000
2500

15

Score

5
0
10

100
2000

3500

Space Invaders

800

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

Score

150

10

Score

4000

200

4000

Score

Score

Score

6000

Q*bert

4500

15

Score

300
250

Pong

20

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

Score

8000

Breakout

350

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

Score

Beamrider

10000

6
8
10
Training time (hours)
Beamrider

12

14

350

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

10000

250
Score

4000

6
8
10
Training time (hours)
Beamrider

12

Score

4000

16000

6
8
10
Training time (hours)
Beamrider

12

6
8
10
Training time (hours)
Breakout

12

12000

Score

6000

6
8
10
Training time (hours)
Breakout

12

12

3000
2500

6
8
10
Training time (hours)
Q*bert

600

12

6
8
10
Training time (hours)
Pong

12

12000

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

6
8
10
Training time (hours)
Q*bert

12

12

14

6
8
10
Training time (hours)
Space Invaders

12

14

6
8
10
Training time (hours)

12

14

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

1400
1200
1000

6000

800
600

4000

4000

400

2000
0

6
8
10
Training time (hours)
Space Invaders

500

1600

8000

10

400

100

14

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

10000

200

500
0

14

300

1000

14

12

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

600

1500
n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

700

2000

6
8
10
Training time (hours)
Space Invaders

500

800

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

400

100

14

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

700

300

800

200

10

200

100

14

500

3500

20

300

12

1000

10

30

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

6
8
10
Training time (hours)
Q*bert

1500

4500

2000

14

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

2500

4000

25

14

15

20
2

300

3000

6
8
10
Training time (hours)
Pong

15

Score

Score
0

500
400

200

3500

10

400

8000

14

20

150

500

10000

12

25

14

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

600

A3C, 1 threads
A3C, 2 threads
A3C, 4 threads
A3C, 8 threads
A3C, 16 threads

14000

6
8
10
Training time (hours)
Pong

20
0

200

14

15

50
0

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

10

100

2000

10

250

6000

20

500

15

150

300

8000

1000

350

n-step Q, 1 threads
n-step Q, 2 threads
n-step Q, 4 threads
n-step Q, 8 threads
n-step Q, 16 threads

1500

20

200

14

2000

15
25

14

Score

10000

Score

12

50

12000

Score

6
8
10
Training time (hours)
Breakout

100

2000

Score

Score

6000

1-step SARSA, 1 threads

1-step SARSA, 2 threads
1-step SARSA, 4 threads
1-step SARSA, 8 threads
1-step SARSA, 16 threads

300

8000

700

Score

12000

Score

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

600

2500

0
10

50

1000

3000

Space Invaders

800

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

3500

Score

100

Q*bert

4000

Score

3000

150

2000

Pong

5
Score

Score

Score

4000

10

200

5000

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

15

Score

250

6000

20

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

Score

7000

Breakout

300

1-step Q, 1 threads
1-step Q, 2 threads
1-step Q, 4 threads
1-step Q, 8 threads
1-step Q, 16 threads

8000

Score

Beamrider

9000

6
8
10
Training time (hours)

12

14

100

20

2000

30

6
8
10
Training time (hours)

12

14

6
8
10
Training time (hours)

12

14

200
0

6
8
10
Training time (hours)

12

14

Figure 7: Training speed comparison of different numbers of actor-learners for all four
asynchronous methods on five Atari games. The x-axis shows training time in hours while
the y-axis shows the average score. Each curve shows the average of the three best performing agents from a search over 50 random learning rates. All asynchronous methods show
significant speedups from using greater numbers of parallel actor-learners.

20

300

150
100
50
10

900
800
700

1000

10

20
30
40
50
Model Rank
Beamrider
1800
A3C, SGD
A3C, RMSProp
1600
A3C, Shared RMSProp
1400

300
200

10000

100
0

10

20
30
Model Rank

40

50

10

20
30
Model Rank

40

50

800
600
200

10

0
20
30
40
50
Model Rank
Seaquest
4000
A3C, SGD
A3C, RMSProp
3500
A3C, Shared RMSProp
3000

20
30
40
50
Model Rank
Space Invaders
A3C, SGD
A3C, RMSProp
A3C, Shared RMSProp

2000

800

1500

600

1000

400

500

200

10

2500

1000

5000

1000

400

1200
Score

Score

400

1400

2000

15000

500

1600
1200

3000

5000
20
30
40
50
Model Rank
Breakout
25000
A3C, SGD
A3C, RMSProp
A3C, Shared RMSProp
20000

600
Score

10000

Space Invaders
n-step Q, SGD
n-step Q, RMSProp
n-step Q, Shared RMSProp

1800

4000
Score

200

5000

15000
Score

Score

250

20000

Seaquest
n-step Q, SGD
n-step Q, RMSProp
n-step Q, Shared RMSProp

6000

Score

350

Beamrider
n-step Q, SGD
n-step Q, RMSProp
n-step Q, Shared RMSProp

25000

Score

Breakout
n-step Q, SGD
n-step Q, RMSProp
n-step Q, Shared RMSProp

400

10

20
30
Model Rank

40

50

10

20
30
Model Rank

Figure 8: Comparison of three different optimization methods (Momentum SGD, RMSProp, Shared RMSProp) tested using two different algorithms (Async n-step Q and Async
Advantage Actor-Critic) on four different Atari games (Breakout, Beamrider, Seaquest and
Space Invaders). Each curve shows the final scores for 50 experiments sorted in descending
order that covers a search over 50 random initializations and learning rates. The top row
shows results using Async n-step Q algorithm and bottom row shows results with Async
Advantage Actor-Critic. Each individual graph shows results for one of the four games and
three different optimization methods. Shared RMSProp tends to be more robust to different learning rates and random initializations than Momentum SGD and RMSProp without
sharing.
ADADELTA [Zeiler, 2012], RMSProp [Tieleman and Hinton, 2012] and ADAM [Kingma
and Ba, 2014], there is no consensus as to which method is the best. In Figure 8 we compare
three different asynchronous optimization algorithms (Momentum SGD, RMSProp, Shared
RMSProp) combined with two different reinforcement learning methods (Async n-step Q
and Async Advantage Actor-Critic) on four different tasks (Breakout, Beamrider, Seaquest
and Space Invaders). Each curve shows the scores for 50 experiments that correspond
to 50 different random learning rates and initializations. The x-axis shows the rank of
the model after sorting in descending order by final average score and the y-axis shows
the final average score achieved by the corresponding model. In this representation, the
algorithm that performs better would achieve higher maximum rewards on the y-axis and
the algorithm that is most robust would have its slope closest to horizontal, thus maximizing
the area under the curve. RMSProp with shared statistics tends to be more robust than
21

40

50

Beamrider

14000

Breakout

1000

12000

2500
1500

400

Score

Score

6000
4000

2000
Score

600

8000
Score

Space Invaders

3000

800

10000

1000

200

1500
1000

2000

500
0

0
2000

Seaquest

2000

10-3

10-2
Entropy cost

200

500
10-3

10-2
Entropy cost

10-3

10-2
Entropy cost

10-3

10-2
Entropy cost

Figure 9: Scatter plots of final scores achieved by Advantage Actor-Critic on four games
(Breakout, Beamrider, Seaquest, Space Invaders) for 50 different entropy regularization
penalty coefficients, learning rates, and random initializations. On some games using entropy regularization improves performance.

RMSProp with per-thread statistics, which is in turn more robust than Momentum SGD.
Next, we look at the stability and robustness of the asynchronous algorithms. We
trained models on five games (Breakout, Beamrider, Pong, Q*bert, Space Invaders) using
four different algorithms (one-step Q, one-step Sarsa, n-step Q and Advantage Actor-Critic)
using 50 different learning rates and random initializations. Scatter plots of scores are shown
for all algorithms and tasks in Figure 10. There is usually a range of learning rates for each
method and game combination that leads to a high score, indicating that all methods are
quite robust to the choice of learning rate. The fact that there are virtually no points with
scores of 0 in regions with good learning rates indicates that the methods are stable and
do not collapse or diverge once they are learning. Similarly, in Figure 9 we show scatter
plots of scores obtained by training Advantage Actor-Critic for 50 combinations of random
initialization, learning rate and entropy cost on four games. These results show that using
entropy regularization with advantage actor-critic does lead to better scores on some games.

Conclusions and Discussion

We have presented asynchronous versions of four standard reinforcement learning algorithms

and showed that they are able to train neural network controllers on a variety of domains
in a stable manner. Our results show that in our proposed framework stable training
of neural networks through reinforcement learning is possible with both value-based and
policy-based methods, off-policy as well as on-policy methods, and in discrete as well as
continuous domains. When trained on the Atari domain using 16 CPU cores, the proposed
asynchronous algorithms train faster than DQN trained on an Nvidia K40 GPU, with A3C
surpassing the current state-of-the-art in half the training time.
One of our main findings is that using parallel actor-learners to update a shared model
had a stabilizing effect on the learning process of the three value-based methods we considered. While this shows that stable online Q-learning is possible without experience replay,
which was used for this purpose in DQN, it does not mean that experience replay is not

22

0
10-4

14000

10-3
Learning rate
1-step SARSA, Beamrider

50

10-2

400

30

10-2

300

10

250

200

Score

10000

6000

150
100
50

15

20

50

10-2

400
350

12000

300

10000

250

8000

200

Score

14000

6000

100

2000

50

25

10-2

10-3
Learning rate
A3C, Beamrider

50

10-2

10-4

1000

14000

10-3
Learning rate
A3C, Breakout

600
Score

8000
6000
4000
2000
0
10-4

10-3
Learning rate

10-2

10-4

10-3
Learning rate

10-2

10-2

1000

10-4

10-3
Learning rate
n-step Q, Space Invaders

10-2

10-4

10-3
Learning rate
A3C, Space Invaders

10-2

10-4

10-3
Learning rate

10-2

600
500
400

20

0
10-3
Learning rate
A3C, Pong

10-4

10-3
Learning rate

10-4

10-2

1000

100

10-2

1000
900
800
700
600
500
400

10-4

12000

10-2

10-3
Learning rate
n-step Q, Q*bert

2000
1000

10-4

200

5000

10

30

Score

Score
10-3
Learning rate
n-step Q, Pong

20

200

10-4

3000

10-2

300

10

10

10-3
Learning rate
1-step SARSA, Space Invaders

800

2000

4000

200

10-4

900

10

400

100

10-2

1000

20

Score

10000

10-3
Learning rate
1-step SARSA, Q*bert

700

30

800

12000

10-4

5000

20

30

10-2

1000

300

3000

30

0
10-4

16000

2000

10-3
Learning rate
n-step Q, Breakout

150

4000

2000

10-4

Score

16000

10-3
Learning rate
n-step Q, Beamrider

10-2

500
400

200

4000

0
10-4

10-3
Learning rate
1-step SARSA, Pong

10

2000

2000

10-4

20
15

Score

Score

350

4000

Score

10-3
Learning rate
1-step SARSA, Breakout

20

12000

8000

Score

10-4

1000

Score

2000

10

Score

50

2000

10-3
Learning rate
A3C, Q*bert

300

10-2

1400

10000

1200

8000

1000

6000

800

Score

100

Score

4000

150

700
600

3000

Score

6000

1-step Q, Space Invaders

800

4000

10

200

Score

Score

Score

250

1-step Q, Q*bert

5000

20

300

8000

1-step Q, Pong

30

350

10000

1-step Q, Breakout

400

Score

1-step Q, Beamrider

12000

4000

600

2000

400

200

2000

10-4

10-3
Learning rate

10-2

Figure 10: Scatter plots of scores obtained by four different algorithms (one-step Q, onestep Sarsa, n-step Q and Advantage Actor-Critic) on five games (Beamrider, Breakout,
Pong, Q*bert, Space Invaders) for 50 different learning rates and random initializations.
All algorithms exhibit some level of robustness to the choice of learning rate.
useful. Incorporating experience replay into the asynchronous reinforcement learning framework could substantially improve the data efficiency of these methods by reusing old data.
This could in turn lead to much faster training times in domains like TORCS where interacting with the environment is more expensive than updating the model for the architecture
we used.
Combining other existing reinforcement learning methods or recent advances in deep
reinforcement learning with our asynchronous framework presents many possibilities for
immediate improvements to the methods we presented. While our n-step methods operate
23

in the forward view [Sutton and Barto, 1998] by using corrected n-step returns directly as
targets, it has been more common to use the backward view to implicitly combine different returns through eligibility traces [Watkins, 1989, Sutton and Barto, 1998, Peng and
Williams, 1996]. The asynchronous advantage actor-critic method could be potentially improved by using other ways of estimating the advantage function, such as generalized advantage estimation of Schulman et al. [2015b]. All of the value-based methods we investigated
could benefit from different ways of reducing over-estimation bias of Q-values [Van Hasselt
et al., 2015, Bellemare et al., 2016]. Yet another, more speculative, direction is to try and
combine the recent work on true online temporal difference methods [van Seijen et al., 2015]
with nonlinear function approximation.
In addition to these algorithmic improvements, a number of complementary improvements to the neural network architecture are possible. The dueling architecture of Wang
et al. [2015] has been shown to produce more accurate estimates of Q-values by including
separate streams for the state value and advantage in the network. The spatial softmax
proposed by Levine et al. [2015] could improve both value-based and policy-based methods
by making it easier for the network to represent feature coordinates.
Acknowledgments
We thank Thomas Degris, Remi Munos, Marc Lanctot, Sasha Vezhnevets and Joseph Modayil for many helpful discussions, suggestions and comments on the paper. We also thank
the DeepMind evaluation team for setting up the environments used to evaluate the agents
in the paper.

24

Game
Alien
Amidar
Assault
Asterix
Asteroids
Atlantis
Bank Heist
Battle Zone
Beam Rider
Berzerk
Bowling
Boxing
Breakout
Centipede
Chopper Comman
Crazy Climber
Defender
Demon Attack
Double Dunk
Enduro
Fishing Derby
Freeway
Frostbite
Gopher
Gravitar
H.E.R.O.
Ice Hockey
James Bond
Kangaroo
Krull
Kung-Fu Master
Montezumas Revenge
Ms. Pacman
Name This Game
Phoenix
Pit Fall
Pong
Private Eye
Q*Bert
River Raid
Road Runner
Robotank
Seaquest
Skiing
Solaris
Space Invaders
Star Gunner
Surround
Tennis
Time Pilot
Tutankham
Up and Down
Venture
Video Pinball
Wizard of Wor
Yars Revenge
Zaxxon

DQN
570.2
133.4
3332.3
124.5
697.1
76108.0
176.3
17560.0
8672.4

Gorila
813.5
189.2
1195.8
3324.7
933.6
629166.5
399.4
19938.0
3822.1

41.2
25.8
303.9
3773.1
3046.0
50992.0

54.0
74.2
313.0
6296.9
3191.8
65451.0

12835.2
-21.6
475.6
-2.3
25.8
157.4
2731.8
216.5
12952.5
-3.8
348.5
2696.0
3864.0
11875.0
50.0
763.5
5439.9

14880.1
-11.3
71.0
4.6
10.2
426.6
4373.0
538.4
8963.4
-1.7
444.0
1431.0
6363.1
20620.0
84.0
1263.0
9238.5

16.2
298.2
4589.8
4065.3
9264.0
58.5
2793.9

16.7
2598.6
7089.8
5310.3
43079.8
61.8
10145.9

1449.7
34081.0

1183.3
14919.2

-2.3
5640.0
32.4
3311.3
54.0
20228.1
246.0

-0.7
8267.8
118.5
8747.7
523.4
112093.4
10431.0

831.0

6159.4

Double
1033.4
169.1
6060.8
16837.0
1193.2
319688.0
886.0
24740.0
17417.2
1011.1
69.6
73.5
368.9
3853.5
3495.0
113782.0
27510.0
69803.4
-0.3
1216.6
3.2
28.8
1448.1
15253.0
200.5
14892.5
-2.5
573.0
11204.0
6796.1
30207.0
42.0
1241.3
8960.3
12366.5
-186.7
19.1
-575.5
11020.8
10838.4
43156.0
59.1
14498.0
-11490.4
810.0
2628.7
58365.0
1.9
-7.8
6608.0
92.2
19086.9
21.0
367823.7
6201.0
6270.6
8593.0

Dueling
1486.5
172.7
3994.8
15840.0
2035.4
445360.0
1129.3
31320.0
14591.3
910.6
65.7
77.3
411.6
4881.0
3784.0
124566.0
33996.0
56322.8
-0.8
2077.4
-4.1
0.2
2332.4
20051.4
297.0
15207.9
-1.3
835.5
10334.0
8051.6
24288.0
22.0
2250.6
11185.1
20410.5
-46.9
18.8
292.6
14175.8
16569.4
58549.0
62.0
37361.6
-11928.0
1768.4
5993.1
90804.0
4.0
4.4
6601.0
48.0
24759.2
200.0
110976.2
7054.0
25976.5
10164.0

Prioritized
900.5
218.4
7748.5
31907.5
1654.0
593642.0
816.8
29100.0
26172.7
1165.6
65.8
68.6
371.6
3421.9
6604.0
131086.0
21093.5
73185.8
2.7
1884.4
9.2
27.9
2930.2
57783.8
218.0
20506.4
-1.0
3511.5
10241.0
7406.5
31244.0
13.0
1824.6
11836.1
27430.1
-14.8
18.9
179.0
11277.0
18184.4
56990.0
55.4
39096.7
-10852.8
2238.2
9063.0
51959.0
-0.9
-2.0
7448.0
33.6
29443.7
244.0
374886.9
7451.0
5965.1
9501.0

A3C FF*
182.1
283.9
3746.1
6723.0
3009.4
772392.0
946.0
11340.0
13235.9
1433.4
36.2
33.7
551.6
3306.5
4669.0
101624.0
36242.5
84997.5
0.1
-82.2
13.6
0.1
180.1
8442.8
269.5
28765.8
-4.7
351.5
106.0
8066.6
3046.0
53.0
594.4
5614.0
28181.8
-123.0
11.4
194.4
13752.3
10001.2
31769.0
2.3
2300.2
-13700.0
1884.8
2214.7
64393.0
-9.6
-10.2
5825.0
26.1
54525.4
19.0
185852.6
5278.0
7270.8
2659.0

A3C FF
518.4
263.9
5474.9
22140.5
4474.5
911091.0
970.1
12950.0
22707.9
817.9
35.1
59.8
681.9
3755.8
7021.0
112646.0
56533.0
113308.4
-0.1
-82.5
18.8
0.1
190.5
10022.8
303.5
32464.1
-2.8
541.0
94.0
5560.0
28819.0
67.0
653.7
10476.1
52894.1
-78.5
5.6
206.9
15148.8
12201.8
34216.0
32.8
2355.4
-10911.1
1956.0
15730.5
138218.0
-9.7
-6.3
12679.0
156.3
74705.7
23.0
331628.1
17244.0
7157.5
24622.0

Table 3: Raw scores for the human start condition (30 minutes emulator time). DQN
scores taken from Nair et al. [2015]. Double DQN scores taken from Van Hasselt et al.
[2015], Dueling scores from Wang et al. [2015] and Prioritized scores taken from Schaul
et al. [2015]
25

A3C LSTM
945.3
173.0
14497.9
17244.5
5093.1
875822.0
932.8
20760.0
24622.2
862.2
41.8
37.3
766.8
1997.0
10150.0
138518.0
233021.5
115201.9
0.1
-82.5
22.6
0.1
197.6
17106.8
320.0
28889.5
-1.7
613.0
125.0
5911.4
40835.0
41.0
850.7
12093.7
74786.7
-135.7
10.7
421.1
21307.5
6591.9
73949.0
2.6
1326.1
-14863.8
1936.4
23846.0
164766.0
-8.3
-6.4
27202.0
144.2
105728.7
25.0
470310.5
18082.0
5615.5
23519.0

References
Marc G Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning
environment: An evaluation platform for general agents. Journal of Artificial Intelligence
Research, 2012.
Marc G. Bellemare, Georg Ostrovski, Arthur Guez, Philip S. Thomas, and Remi Munos.
Increasing the action gap: New operators for reinforcement learning. In Proceedings of
the AAAI Conference on Artificial Intelligence, 2016.
Dimitri P Bertsekas. Distributed dynamic programming. Automatic Control, IEEE Transactions on, 27(3):610616, 1982.
Kevin Chavez, Hao Yi Ong, and Augustus Hong. Distributed deep q-learning. Technical
report, Stanford University, June 2015.
Thomas Degris, Patrick M Pilarski, and Richard S Sutton. Model-free reinforcement learning with continuous action in practice. In American Control Conference (ACC), 2012,
pages 21772182. IEEE, 2012.
John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online
learning and stochastic optimization. The Journal of Machine Learning Research, 12:
21212159, 2011.
Matthew Grounds and Daniel Kudenko. Parallel reinforcement learning with linear function
approximation. In Proceedings of the 5th, 6th and 7th European Conference on Adaptive
and Learning Agents and Multi-agent Systems: Adaptation and Multi-agent Learning,
pages 6074. Springer-Verlag, 2008.
Sepp Hochreiter and J
urgen Schmidhuber. Long short-term memory. Neural computation,
9(8):17351780, 1997.
Tommi Jaakkola, Michael I Jordan, and Satinder P Singh. On the convergence of stochastic
iterative dynamic programming algorithms. Neural computation, 6(6):11851201, 1994.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv
preprint arXiv:1412.6980, 2014.
Jan Koutnk, J
urgen Schmidhuber, and Faustino Gomez. Evolving deep unsupervised convolutional networks for vision-based reinforcement learning. In Proceedings of the 2014
conference on Genetic and evolutionary computation, pages 541548. ACM, 2014.
Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of
deep visuomotor policies. arXiv preprint arXiv:1504.00702, 2015.
Yuxi Li and Dale Schuurmans. Mapreduce for parallel reinforcement learning. In Recent
Advances in Reinforcement Learning - 9th European Workshop, EWRL 2011, Athens,
Greece, September 9-11, 2011, Revised Selected Papers, pages 309320, 2011.

26

Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval
Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement
learning. arXiv preprint arXiv:1509.02971, 2015.
Long-Ji Lin. Reinforcement learning for robots using neural networks. Technical report,
DTIC Document, 1993.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou,
Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning.
In NIPS Deep Learning Workshop. 2013.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G.
Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski,
Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529533, 02 2015. URL
http://dx.doi.org/10.1038/nature14236.
Arun Nair, Praveen Srinivasan, Sam Blackwell, Cagdas Alcicek, Rory Fearon, Alessandro De
Maria, Vedavyas Panneershelvam, Mustafa Suleyman, Charles Beattie, Stig Petersen,
Shane Legg, Volodymyr Mnih, Koray Kavukcuoglu, and David Silver. Massively parallel
methods for deep reinforcement learning. In ICML Deep Learning Workshop. 2015.
Jing Peng and Ronald J Williams. Incremental multi-step q-learning. Machine Learning,
22(1-3):283290, 1996.
Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free
approach to parallelizing stochastic gradient descent. In Advances in Neural Information
Processing Systems, pages 693701, 2011.
Martin Riedmiller. Neural fitted q iterationfirst experiences with a data efficient neural reinforcement learning method. In Machine Learning: ECML 2005, pages 317328.
Springer Berlin Heidelberg, 2005.
Gavin A Rummery and Mahesan Niranjan. On-line q-learning using connectionist systems.
1994.
Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience
replay. arXiv preprint arXiv:1511.05952, 2015.
John Schulman, Sergey Levine, Philipp Moritz, Michael I Jordan, and Pieter Abbeel. Trust
region policy optimization. In International Conference on Machine Learning (ICML),
2015a.
John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. arXiv preprint
arXiv:1506.02438, 2015b.

27

Satinder Singh, Tommi Jaakkola, Michael L Littman, and Csaba Szepesvari. Convergence
results for single-step on-policy reinforcement-learning algorithms. Machine Learning, 38
(3):287308, 2000.
R. Sutton and A. Barto. Reinforcement Learning: an Introduction. MIT Press, 1998.
Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a
running average of its recent magnitude. COURSERA: Neural Networks for Machine
Learning, 4, 2012.
E Todorov. MuJoCo: Modeling, Simulation and Visualization of Multi-Joint Dynamics
with Contact (ed 1.0). Roboti Publishing, 2015.
Marco Tomassini. Parallel and distributed evolutionary algorithms: A review. Technical
report, 1999.
J. Tsitsiklis and B. Van Roy. An analysis of temporal-difference learning with function
approximation. IEEE Transactions on Automatic Control, 42(5):674690, 1997.
John N Tsitsiklis. Asynchronous stochastic approximation and q-learning. Machine Learning, 16(3):185202, 1994.
Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double
q-learning. arXiv preprint arXiv:1509.06461, 2015.
H. van Seijen, A. Rupam Mahmood, P. M. Pilarski, M. C. Machado, and R. S. Sutton. True
Online Temporal-Difference Learning. ArXiv e-prints, December 2015.
Z. Wang, N. de Freitas, and M. Lanctot. Dueling Network Architectures for Deep Reinforcement Learning. ArXiv e-prints, November 2015.
Christopher John Cornish Hellaby Watkins. Learning from delayed rewards. PhD thesis,
University of Cambridge England, 1989.
R.J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3):229256, 1992.
Ronald J Williams and Jing Peng. Function optimization using connectionist reinforcement
learning algorithms. Connection Science, 3(3):241268, 1991.
B. Wymann, E. Espi, C. Guionneau, C. Dimitrakakis, R. Coulom, and A. Sumner. Torcs:
The open racing car simulator, v1.3.5, 2013.
Matthew D Zeiler. Adadelta:
arXiv:1212.5701, 2012.

An adaptive learning rate method.

28

arXiv preprint