1Managing the 802.11 Energy/Performance
Tradeoff with Machine Learning
Claire Monteleoni Hari Balakrishnan Nick Feamster Tommi Jaakkola
MIT Computer Science and Artificial Intelligence Laboratory
32 Vassar St.
Cambridge, MA 02139
{cmontel, hari, feamster, tommi}@csail.mit.edu
Abstract—This paper addresses the problem of managing
the tradeoff between energy consumption and performance
in wireless devices implementing the IEEE 802.11 standard
[1]. To save energy, the 802.11 specification proposes a
power-saving mode (PSM), where a device can sleep to save
energy, periodically waking up to receive packets from a
neighbor (e.g., an access point) that may have buffered pack-
ets for the sleeping device. Previous work has shown that a
fixed polling time for waking up degrades the performance
of Web transfers [2], because network activity is bursty and
time-varying. We apply a new online machine learning algo-
rithm to this problem and show, using ns-2 simulation and
trace analysis, that it is able to adapt well to network activity.
The learning process makes no assumptions about the un-
derlying network activity being stationary or even Markov.
Our learning power-saving algorithm, LPSM, guides the
learning using a “loss function” that combines the increased
latency from potentially sleeping too long and the wasted use
of energy in waking up too soon. In our ns-2 simulations,
LPSM saved 7%-20% more energy than 802.11 in power-
saving mode, with an associated increase in average latency
by a factor of 1.02, and not more than 1.2. LPSM is straight-
forward to implement within the 802.11 PSM framework.
I. INTRODUCTION
Energy is an important resource in mobile computing
systems. Because processing, storage, display activity,
and communication all consume energy, energy-saving
techniques targeted at improving these subsystems have
received significant attention in recent years. Impressive
advances in hardware design and operating systems have
greatly reduced the energy consumed by the processing
and storage subsystems, and have led to the wireless net-
work becoming a significant consumer of energy in many
mobile devices. This trend is especially true for handheld
mobile devices and nodes in wireless ad hoc and sensor
networks.
Most wireless network interfaces consume energy not
only while transmitting or receiving data, but also when
they are simply awake. Therefore, to save energy, most
modern wireless interfaces support a power saving mode
(PSM). In abstract terms, the PSM primitive allows an in-
terface to be in one of two states, SLEEP or AWAKE.
SLEEP is a low-power state, but the interface cannot
send or receive data in this state. In contrast, AWAKE
allows data flow, but is a higher-power state. Depending
on the actual device, these two states may differ in power
consumption by between a factor of 10 and 50. For in-
stance, in some current 802.11 cards, the ratio is about a
factor of 20 (1 W v. 50 mW) [3], [4].
With the PSM primitive, power-saving algorithms can
save energy by keeping the wireless interface in the
SLEEP state for as long as possible. A SLEEPing de-
vice periodically wakes up and polls its neighbors (either
an access point in “infrastructure” mode, or a neighboring
node in “ad hoc” mode) for packets.1 To avoid excessive
packet loss, the neighbor must therefore buffer packets for
each SLEEPing receiver. Then, the neighbor sends these
buffered packets when it receives a poll from a waking
receiver.
Power-saving algorithms built on top of the PSM prim-
itive introduce a tradeoff between the amount of energy
saved and the degree of performance degradation. If a
device awakens and finds no data buffered for it, then it
could have slept for longer and saved some more energy.
On the other hand, if any packets are buffered when the
interface awakens, then the latency to obtain those pack-
ets would be larger than if the network interface had been
awake instead of asleep. This increased latency degrades
not just the latency of the on-going data transfers, but of-
ten the throughput as well.
This paper addresses the problem designing an algo-
rithm by which a device can decide when to SLEEP and
when to be AWAKE. Our goal is to devise an algorithm
that manages the tradeoff between energy consumption
and data transfer latency in a principled, well-specified
way, such that users or application designers can specify
1This is an abstract model: some implementations first have the neigh-
bor advertise information before the polls occur.

2their desired operating point. Our motivation for a prin-
cipled trade-off is motivated by Krashinsky and Balakr-
ishnan’s work on the Bounded SlowDown (BSD) algo-
rithm [2], where they demonstrate that the IEEE 802.11’s
non-adaptive polling time strategy [1] degrades both the
latency and the throughput of TCP transfers. As in their
work, we focus our algorithm on Web-like workloads,
mainly because it is the dominant workload today for
many mobile devices.
We develop a PSM algorithm called LPSM (Learn-
ing PSM) to determine a device’s sleep/awake schedule.
LPSM adapts the schedule to network activity by main-
taining a bank of “experts”. Each expert is a determin-
istic setting of the polling time, with an associated time-
varying weight that gets updated by the algorithm in re-
sponse to observations of current network activity. The
weights form a probability distribution, and at each time,
the polling time chosen is the weighted sum of the ex-
perts’ times. LPSM is based on a recently developed ma-
chine learning algorithm called Learn-α [5], [6], which
has the attractive property that it makes no assumptions
on the statistical distribution of the activity being learned.
In our context, LPSM’s use of Learn-α makes no assump-
tions on the distribution of packet arrivals and network
activity.
The first contribution of this paper is to show how on-
line machine learning can be used to solve the wireless
power-saving problem. The key to this solution is to de-
fine a loss function that the Learn-α algorithm uses in de-
termining how to update the weights of the experts every
time the mobile device awakens. If the device awakens
and there is no data present, then the weights of the ex-
perts are carefully adjusted such that the next sleep time
is longer. Conversely, if any packets were present, the
opposite adjustment is made.
The second contribution of this paper is a performance
evaluation of LPSM using trace-driven simulations. We
compare the performance of both the non-adaptive 802.11
PSM and the BSD algorithm to LPSM. Our experimen-
tal results to date have shown that for a Web-like re-
quest/response workload, LPSM saves 7%-20% more en-
ergy than 802.11 in power-saving mode, with an associ-
ated increase in average slowdown of 2%, and not more
than 20%. LPSM is straightforward to implement within
the 802.11 PSM framework.
The rest of this paper is organized as follows. In Sec-
tion II, we survey related work in power saving and ma-
chine learning. Section III gives a formal definition of
the problem and an overview of our approach. Section IV
gives the LPSM algorithm. Section V presents several
results from trace-driven ns-2 simulations and trace-
based analysis of LPSM, and Section VI concludes the
paper with a discussion of our results.
II. RELATED WORK
Using trace-driven simulations, Krashinsky and Bal-
akrishnan [2] show that the 802.11 PSM algorithm, which
uses a fixed polling interval (typically 100 ms) to wake
up and check for data, causes response latency for Web-
like transfers to be as bad as 2.2× longer than in the ab-
sence of any power-saving algorithm. To better manage
the tradeoff in question, they proposed BSD, an algorithm
that uses an adaptive control loop to change polling time
based on network conditions. The algorithm uses a pa-
rameter, p, and guarantees that the response latency does
not ever exceed (1 + p) times the response latency with-
out power-saving. Within that constraint and assuming
adversarial traffic arrivals, BSD guarantees that the en-
ergy consumption is minimized. In contrast, LPSM does
not attempt to guarantee bounded latency under adversar-
ial traffic arrivals; instead, our approach is to explicitly
encode a tradeoff between energy and latency and give an
online learning algorithm that manages this tradeoff.
Simunic et al. formulate the wireless power-saving
problem as policy learning in a Markov Decision Process
(MDP) [7]. Their algorithm is not an online algorithm
since the linear programming algorithm used to resolve
the policy over any given time period requires access to
data over that entire period. They also assume that net-
work activity is stationary. In the MDP there is fixed dis-
tribution governing the selection of next states, given the
current state and action. For any fixed policy such as the
optimal policy in this framework, the network activity is
modeled as a Markov process. This model is not an ideal
one for a mobile node, since the network activity need not
conform to a Markov process of any finite order k.
Simunic et al. refer to Chung et al. [8] for the solu-
tion in non-stationary environments. That work proposes
“policy interpolation”; however, it still assumes that the
underlying process is Markovian, even though it may ini-
tially appear non-stationary due to a lack of observations.
They then propose to learn the associated MDP parame-
ters sequentially [8]. Another machine learning approach
to this problem was proposed Steinbach, using Reinforce-
ment Learning [9]. This approach also imposes the as-
sumption that network activity has the Markov property
which, as discussed above, is unrealistic.
These previous learning approaches differ from ours in
that LPSM does not make any Markovian or stationarity
assumptions, nor require any a priori knowledge of the
statistical process being learned. LPSM is also simpler
to implement in the 802.11 framework compared to these
previous learning approaches.
Our learning algorithm is from the field of online learn-
ing algorithms. Littlestone and Warmuth [10] designed
online learning algorithms that have access to a fixed set
of experts, for which they were able to bound the cumula-