Content uploaded by Michael C. Mozer
Author content
All content in this area was uploaded by Michael C. Mozer on Jul 29, 2015
Content may be subject to copyright.
The Neural Network House:
An Environment that Adapts to its Inhabitants
Michael C. Mozer
Department of Computer Science and
Institute of Cognitive Science
University of Colorado
Boulder, CO 80309-0430
mozer@colorado, edu
Abstract
Although the prospect of computerized homes has a long
history, ho/ne automation has never become terribly popular
because the benefits are seldom seen to outweigh the costs.
One significant cost of an automated home is that someone
has to program it to behave appropriately. Typical inhabit-
ants do not want to program simple devices such as VCRs,
let alone a much broader range of electronic devices, appli-
ances, and comfort systems that have even greater function-
ality. We describe an alternative approach t in which the goal
is for the home to essentially program itself by observing the
lifestyle and desires of the inhabitants, and learning to antic-
ipate and accommodate their needs. The system we have
developed controls basic residential comfort systems--air
heating, lighting, ventilation, and water heating. We have
constructed a prototype system in an actual residence, and
describe initial results and the current state of the project.
Introduction
Since the mid 1940s, the home automation industry has
promised to revolutionize our living environments. The so-
called "smart home" has been hyped in the popular press.
The vision of the. industry is that household devices--appli-
ances, entertainment centers, utilities, thermostats, lights,
etc.--will be endowed with microprocessors that allow the
devices to communicate with one another and thereby
behave intelligently. The dishwasher can ask the hot water
heater whether it has sufficient capacity to operate; inhabit-
ants can telephone home and remotely instruct the VCR to
record a favorite show; the TV might lower its volume
when the phone rings; or the clothes dryer might make an
announcement over an intercom system when it has com-
pleted its cycle.
As attractive as this scenario is, the software required to
achieve the intelligence is highly complex and unwieldy,
and worse, the software must be tailored to a particular
home and family, and updated as the family’s lifestyle
changes. Tackling the programming task is far beyond the
capabilities and interest of typical home inhabitants.
Indeed, even rudimentary forms of regulation, such as oper-
ating a set back thermostat, which allows different tempera-
t.ure settings depending on the time of day, are inordinately
difficult for people (Gregorek, 1991). The alternative of hir-
ing professional technicians to update programs as neces-
sary is used in some commercial systems, but is costly and
inconvenient. Partly due to these difficulties in program-
ming, home automation has never become a widely avail-
able and accepted technology.
In contrast to standard computerized homes that can be
programmed to perform various functions, the crux of our
project is to develop a home that essentially programs itself
by observing the lifestyle and desires of the inhabitants, and
learning to anticipate and accommodate their needs. The
system we have developed controls basic residential com-
fort systems--air heating, lighting, ventilation, and water
heating.
ACHE
we call the system ACHE, which stands for adaptive cor/-
trol of home _environments. ACHE monitors the environ-
ment, observes the actions taken by occupants (e.g.,
adjusting the thermostat; turning on a particular configura-
tion of lights), and attempts to infer patterns in the environ-
ment that predict these actions.
ACHE has two objectives. One is anticipation of inhabit-
ants’ needs. Lighting, air temperature, and ventilation
should be maintained to the inhabitants’ comfort; hot water
should be available on demand. When inhabitants manually
adjust environmental setpoints, it is an indication that their
needs have not been satisfied and will serve as a training
signal for ACHE. If ACHE can learn to anticipate needs,
manual control of the environment will be avoided. The
second objective of ACHE is energy conservation. Lights
should be set to the minimum intensity required; hot water
should be maintained at the minimum temperature needed
to satisfy the demand; only rooms that are likely to be occu-
II0
From: AAAI Technical Report SS-98-02. Compilation copyright © 1998, AAAI (www.aaai.org). All rights reserved.
Figure 1. The Neural Network House, circa 1926.
pied in the near future should be heated; when several
options exist to heat a room (e.g., furnace, ceiling fans forc-
ing hot air down, opening blinds to admit sunlight), the
alternative minimizing expected energy consumption
¯ should be selected.
Achieving either one of these objectives in isolation is
fairly straightforward. If ACHE were concerned only with
appeasing the inhabitants, the air temperature could be
maintained at a comfortable 70
°
at all times. If ACHE were
concerned only with energy conservation, all devices could
be turned off. ACHE’s challenge is to achieve both objec-
tives simultaneously. This requires the ability to anticipate
inhabitant activities, occupancy patterns, and tolerances.
Optimal Control
In what sort of framework can the two objectives--appeas-
ing the inhabitants and conserving energy----be integrated?
Supervised learning will not do: If a temperature setpoint
chosen by the inhabitant serve as the target for a supervised
learning system, energy costs will not be considered.
Instead, we have adopted an optimal control framework in
which failing to satisfy each objective has an associated
cost. A discomfort cost is incurred if inhabitant preferences
are not met, i.e., if the inhabitant is not happy with the set-
tings determined by ACHE, as indicated by manual control
of the environment. An energy cost is incurred based on the
use of electricity or gas resources. The expected average
cost, J(t0), starting at time
o
can then be expressed as
I
to+K
]
J(to) = E lim 1
[.~¢._~ ,~ E d(xt)
+
e(ut)
t=to +1
where d(xt) is the discomfort cost associated with the
environmental state x at time t, and e(ut) is the energy cost
associated with the control decision u at time t. The goal is
to find an optimal control policy--a mapping from states x
t
to decisions u
t-that
minimizes the expected average cost.
This framework requires that discomfort and energy
costs be expressed in the same currency. We have chosen
dollars as this currency, which makes a characterization of
energy costs straightforward. Relative discomfort is indi-
cated by overriding the choices of ACHE, and this relative
discomfort is translated to a dollar amount by means of a
misery-to-dollars conversion factor. One technique we have
explored for determining this factor, based on an economic
analysis, depends on the loss in productivity that occurs
when ACHE ignores the inhabitants’ desires. Another tech-
nique adjusts the conversion factor over a several month
period based on how much inhabitants are willing to pay for
gas and electricity.
Implementation
We have implemented ACHE in an actual residence. The
residence is a former three-room school house built in 1905
near Boulder, Colorado, originally serving children of the
mining town of Marshall (Figure 1). The school was closed
in 1956 and was completely renovated in 1992, at which
time the infrastructure needed for the ACHE project was
incorporated into the house, including nearly five miles of
low-voltage conductor for collecting sensor data and a
power-line communication system for controlling lighting,
fans, and electric outlets. The residence is an ideal candi-
date for intelligent energy management because of its age,
13-25 foot ceilings, and exposed south and west faces that
hold potential for passive solar heating.
ACHE is equipped with sensors that report the state of
the environment. The sensory state includes the following
for each room in the home:
¯ status of lights (on or off, and if on, intensity level)
¯ status of fans (speed)
¯
status of temperature control user interface (a fancy
digital thermostat that specifies the current setpoint
Iii
r+ + +
3
.................. __I
! ,...,, u n ; I Fl;i
Q
¯ i Q
"-’1- ................... i_ .........
Figure 2. A floor plan of the adaptive house, including locations of sensors and actuators.
temperature for the room, and can be adjusted by the
inhabitant)
¯ ambient illumination
¯ room temperature
¯ sound level
¯
motion detector activity (motion or no motion)
¯
status of all doors and windows (open or closed).
In addition, the system receives the following global infor-
mation:
¯ water heater temperature
¯ water heater energy usage
¯ water heater outflow
¯ furnace energy usage
¯ outdoor temperature
¯ outdoor insolation (sunlight)
¯ gas and electricity costs
¯
time of day, day of week, date.
At present, ACHE has the ability to control the following
actuators:
¯ on/off status and intensity of light banks (22 total)
¯ on/off status and speed of ceiling fans (6 total)
¯ on/off status of water heater
¯ on/off status of gas furnace
¯
on/off status of electric space heaters (2 total)
¯
on/off status of speakers in each room through which
computer can communicate (12 total)
Figure 2 shows a floor plan of the residence, as well as the
approximate location of selected sensors and actuators.
ACHE Architecture
Adaptive control of building energy systems is difficult. We
have incomplete models of the ehvironment and controlled
devices. The environment, including the behavior of the
inhabitants, is nonstationary and stochastic. Controlled
devices are nonlinear. Multiple interacting devices must be
controlled simultaneously. Under such circumstances, tradi-
tional techniques from control theory and artificial intelli-
gence have great difficulty (Dean & Wellman, 1991).
The basic system architecture of ACHE is presented in
Figure 3. This architecture is replicated for each control
domain--lighting, air heating, water heating, and ventila-
tion. The instantaneous environmental state is fed through a
state transformation that computes statistics such as aver-
ages, minima, maxima, and variances in a given temporal
window. The result is a state representation that provides
more information about the environment than the instanta-
neous values. The instantaneous state is also given to an
occupancy model that determines for each zone of the
house--usually corresponding to a room--whether or not it
is occupied. The occupancy model relies on motion detec-
tor signals, but it includes rules that say, essentially, "a zone
remains occupied, even when there is no motion, unless
there is motion in an adjacent zone that was previous unoc-
cupied." Consequently, the occupancy model maintains
occupancy status even when there is no motion.
The three adaptive components of ACHE are shown in
the top of Figure 3. Various predictors attempt to take the
current state and forecast future states. Examples of predic-
tions include: expected occupancy patterns in the house
over the next few hours, expected hot water usage, likeli-
112
~decision
I
device
regulator I
Tsetooint
prdfile
I
setpoint
generator I
l information
ictors
state
?
occupied
representation l I
zones
I
state trans"
moael Iformation I I
°ccupancy
4L
I
instantaneous
environmental state
Figure 3. System architecture of ACHE
hood that a zone will be entered in the next few seconds.
The predictors are implemented as feedforward neural net-
works trained with back propagation, or as a combination
of a neural net and a look up table.
Given the predictions of future states, control decisions
need to be made concerning the energy devices in the home.
The decision making process is split into two stages. The
setpoint generator determines a setpoint profile specifying
the target value of some environmental variable (lighting
level, air temperature, water temperature, etc.) over a win-
dow of time. The device regulator controls physical devices
to achieve the setpoint. The device regulator may have
many alternative devices at its disposal. It must determine
which one or which subset to use.
The reason for dividing control between the setpoint gen-
erator and device regulator is to encapsulate knowledge.
The setpoint generator requires knowledge about inhabitant
preferences, while the device regulator has knowledge
about the physical layout and characteristics of the environ-
ment and controlled devices. If the inhabitants or their pref-
erences change over time, only the setpoint generator need
relearn.
The setpoint generator and device regulator in each
domain are based on one of two approaches to control: indi-
rect control using dynamic programing and models of the
environment and inhabitant, or direct control using rein-
forcement learning. For example, the device regulator for
indoor air temperature uses a predictive model of the indoor
air temperature, as a function of the current indoor tempera-
ture, outdoor temperature, and the states of the furnace and
electric space heaters. This model is based on a simple RC
thermal model of the house and furnace, with a neural net-
work that learns deviations from this simple model and the
actual behavior of the house. Given this model, achieving a
particular setpoint temperature involves little more than
exhaustively searching through the space of heating device
actions and finding an action sequence that achieves the set-
point. In contrast to this indirect approach, the setpoint gen-
erator for the lighting controller uses a direct approach with
reinforcement learning because it would be difficult to learn
an explicit model of inhabitant preferences.
Current Implementation Status
We have conducted simulation studies of the heating con-
trol system (Mozer, Vidmar, & Dodier, 1997), using actual
occupancy data and outdoor temperature profiles, evaluat-
ing various control policies. ACHE robustly outperforms
three alternative policies, showing a lower total (discomfort
plus energy) cost across a range of values for the relative
cost of inhabitant discomfort and the degree of nondeter-
minism in occupancy patterns.
We have also implemented and tested a lighting control-
ler in the house (Mozer & Miller, in press). To give the fla-
vor of its operation, we describe a sample scenario of its
behavior. The first time that the inhabitant enters a zone
(we’ll refer to this as a trial), ACHE decides to leave the
light off, based on the initialization assumption that the
inhabitant has no preference with regard to light settings. If
the inhabitant overrides this decision by turning on the
light, ACHE immediately learns that leaving the light off
will incur a higher cost (the discomfort cost) than turning
on the light to some intensity (the energy cost). On the next
trial, ACHE decides to turn on the light, but has no reason
to believe that one intensity setting will be preferred over
113
another. Consequently, the lowest intensity setting is
selected. On any trial in which the inhabitant adjusts the
light intensity upward, the decision chosen by ACHE will
incur a discomfort cost, and on the following trial, a higher
intensity will be selected. Training thus requires just three
or four trials, and explores the space of decisions to find the
lowest acceptable intensity. ACHE also attempts to con-
serve energy by occasionally "testing" the inhabitant,
selecting an intensity setting lower than the setting believed
to be optimal. If the inhabitant does not complain, the cost
of the decision is updated to reflect this fact, and eventually
the lower setting will be evaluated as optimal.
Evaluating ACHE
It is our conviction that intelligent control techniques for
complex systems in dynamic environments must be devel-
oped and evaluated in naturalistic settings such as the Neu-
ral Network House. While there are numerous examples
illustrating the potential of neural nets for control of build-
ing energy systems (e.g., Curtiss, Kreider, & Brandemuehl,
1994; Miller & Seem, 1991; Seem & Braun, 1991; Scott,
Shavlik, & Ray, 1992), this research focuses on narrowly
defined problems and is generally confined to computer
simulations. The research that does involve control of
actual equipment makes simplifying assumptions about
operating conditions and the environment. We intend to
show that adaptive control will yield benefits in natural
environments under realistic operating conditions.
The research program hinges on a careful evaluation
phase. In the long term, the primary empirical question we
must answer is whether there are sufficiently robust regular-
ities in the inhabitants’ behavior that ACHE can benefit
from them. On first consideration, most people conclude
that their daily schedules are not "regular"; they sometimes
come home at 5 p.m., sometimes at 6 p.m., sometimes not
until 8 p.m. However, even subtle statistical patterns in
behavior--such as the fact that if one is not home at 3 a.m.,
one is unlikely to be home at 4 a.m.--are useful to ACHE.
These are patterns that people are not likely to consider
when they discuss the irregularities of their daily lives.
These patterns are certainly present, and we believe that
they can be usefully exploited in adaptive control of living
environments.
Acknowledgements
We are grateful to Marc Anderson and Robert Dodier, who
helped develop the software infrastructure for the Neural
Network House. The Neural Network House is supported
by the Sensory Home Automation Research Project
(SHARP) of Sensory Inc., as well as a CRCW grant-in-aid
from the University of Colorado, NSF award IRI-9058450,
McDonnell-Pew award 97-18.
References
Curtiss, P. S., Kreider, J. E, & Brandemuehl, M. J. (1994). Local
and global control of commercial building HVAC systems
using artificial neural networks. Proceedings of the 1994
American Control Conference ACC’94. New York: IEEE
Press.
Dean, T. L., & Wellman, M. R (1991). Planning and control. San
Mateo, CA: Morgan Kaufmann.
Gregorek, T. (1991). The energy revolution. Electronic House, 6,
10-15.
Miller, R. C., & Seem, J. E. (1991). Comparison of artificial neural
networks with traditional methods of predicting return time
from night or weekend setback. ASHRAE Transactions, 97.
Mozer, M. C., Dodier, R., & Vidmar, S. (1997). The Neurothermo-
stat: Adaptive control of residential heating systems. In M. C.
Mozer, M. I. Jordan, & T. Petsche (Eds.), Advances in Neural
Information Processing Systems 9 (pp. 953-959). Cambridge,
MA: MIT Press.
Mozer, M. C., & Miller, D. J. (in press). Parsing the stream
time: The value of event-based segmentation in a complex,
real-world control problem. In M. Gori (Ed.), Adaptive pro-
cessing of temporal information. Springer Verlag.
Scott, G. M., Shavlik, J. W., & Ray, W. H. (1992). Refining PID
controllers using neural nets. In J. E. Moody, S. J. Hanson, &
R. P. Lippmann (Eds.), Advances in Neural Information Pro-
cessing Systems 4 (pp 555-562). San Mateo, CA: Morgan
Kaufmann.
Seem, J. E., & Braun, J. E. (1991). Adaptive methods for real-time
forecasting of building electrical demand. ASHRAE Transac-
tions 97.
114