The neural network house: An environment hat adapts to its inhabitants, MC Mozer

Tags: outdoor temperature, energy costs, Acknowledgements Weare, Sensory Inc., Optimal Control, University of Colorado, Relative discomfort, conversion factor, ceiling fans, energy consumption, energy conservation, inhabitant, IEEE Press, software infrastructure, temperature, occupancy model, water heater, room temperature, air temperature, instantaneous state, Robert Dodier, Sensory Home Automation Research Project, Neural Network House, gas furnace, References Curtiss
Content: The Neural Network House: AnEnvironmenthat Adaptsto its Inhabitants Michael C. Mozer Department of Computer Science and Institute of CognitiveScience University of Colorado Boulder, CO 80309-0430 [email protected], edu
From: AAAI technical report SS-98-02. Compilation copyright © 1998, AAAI ( All rights reserved.
Abstract Althoughthe prospect of computerizedhomeshas a long history, ho/neautomationhas neverbecomteerribly popular becausethe benefits are seldomseen to outweighthe costs. Onesignificant cost of an automatedhomeis that someone has to programit to behaveappropriately.Typicalinhabitants do not wantto programsimpledevices such as VCRs, let alonea muchbroaderrangeof electronicdevices,appliances, andcomfortsystemsthat haveevengreater functionality. Wedescribeanalternativeapproacht in whichthe goal is for the homteo essentiallyprogramitself by observingthe lifestyle anddesiresof the inhabitants,andlearningto anticipate and accommodatteheir needs. Thesystem we have developedcontrols basic residential comfortsystems--air heating, lighting, ventilation, andwaterheating. Wehave constructeda prototypesystemin an actual residence, and describeinitial resultsandthe currentstate of theproject. Introduction Since the mid 1940s, the home automation industry has promisedto revolutionize our living environments. The socalled "smart home"has been hyped in the popular press. Thevision of the. industry is that householddevices--appliances, entertainmentcenters, utilities, thermostats, lights, etc.--will be endowedwith microprocessors that allow the devices to communicate with one another and thereby behaveintelligently. The dishwasher can ask the hot water heater whetherit has sufficient capacity to operate; inhabitants can telephone homeand remotely instruct the VCRto record a favorite show; the TVmight lower its volume whenthe phonerings; or the clothes dryer might makean announcementover an intercom system when it has completedits cycle. Asattractive as this scenario is, the softwarerequired to achieve the intelligence is highly complexand unwieldy, and worse, the software must be tailored to a particular homeand family, and updated as the family's lifestyle changes. Tackling the programmingtask is far beyondthe capabilities and interest of typical homeinhabitants.
Indeed, even rudimentaryformsof regulation, such as operating a set back thermostat, whichallows different temperat.ure settings dependingon the time of day, are inordinately difficult for people(Gregorek,1991). Thealternative of hiring professional technicians to update programsas necessary is used in somecommercialsystems, but is costly and inconvenient. Partly due to these difficulties in programming, homeautomation has never becomea widely available and accepted technology. In contrast to standard computerized homesthat can be programmedto perform various functions, the crux of our project is to developa homethat essentially programsitself byobservingthe lifestyle and desires of the inhabitants, and learning to anticipate and accommodatetheir needs. The system we have developed controls basic residential comfort systems--air heating, lighting, ventilation, and water heating. ACHE wecall the system ACHEw, hichstands for adaptive cor/trol of home_environments. ACHEmonitors the environment, observes the actions taken by occupants (e.g., adjusting the thermostat; turning on a particular configuration of lights), and attemptsto infer patterns in the environmentthat predict these actions. ACHhEas two objectives. Oneis anticipation of inhabitants' needs. Lighting, air temperature, and ventilation should be maintainedto the inhabitants' comfort; hot water should be available on demand.Wheninhabitants manually adjust environmentalsetpoints, it is an indication that their needs have not been satisfied and will serve as a training signal for ACHEI.f ACHEcan learn to anticipate needs, manual control of the environment will be avoided. The second objective of ACHEis energy conservation. Lights should be set to the minimumintensity required; hot water should be maintained at the minimumtemperature needed to satisfy the demand;only roomsthat are likely to be occu-
Figure 1. TheNeuralNetworkHouse,circa 1926.
pied in the near future should be heated; whenseveral optionsexist to heat a room(e.g., furnace, ceiling fans forcing hot air down,opening blinds to admit sunlight), the alternative minimizing expected energy consumption Ї shouldbe selected. Achievingeither one of these objectives in isolation is fairly straightforward. If ACHwE ere concerned only with appeasing the inhabitants, the air temperature could be maintained at a comfortable 70° at all times. If ACHwE ere concernedonly with energy conservation, all devices could be turned off. ACHE'cshallenge is to achieve both objectives simultaneously.This requires the ability to anticipate inhabitant activities, occupancypatterns, andtolerances.
optimal control In what sort of frameworkcan the two objectives--appeasing the inhabitants and conservingenergy----be integrated? Supervisedlearning will not do: If a temperature setpoint chosenby the inhabitant serve as the target for a supervised learning system, energy costs will not be considered. Instead, we have adopted an optimal control frameworkin whichfailing 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 happywith the settings determined by ACHEa,s indicated by manualcontrol of the environment.Anenergy cost is incurred based on the use of electricity or gas resources. The expected average cost, J(t0), starting at time o can t hen be expressed as
I to+K J(to) = E lim 1
[.~ў._~ ,~t=Etod+(1xt) + e(ut)
where d(xt) is the discomfort cost associated with the environmentalstate x at time t, ande(ut) is the energycost associated with the control decision u at time t. Thegoal is
to find an optimal control policy--a mappingfromstates xt to decisions ut-that minimizesthe expected average cost. This framework requires that discomfort and energy costs be expressed in the same currency. Wehave chosen dollars as this currency, whichmakesa characterization of energy costs straightforward. Relative discomfort is indicated by overriding the choices of ACHEa,nd this relative discomfort is translated to a dollar amountby meansof a misery-to-dollars conversion factor. Onetechnique wehave explored for determining this factor, based on an economic analysis, depends on the loss in productivity that occurs whenACHiEgnores the inhabitants' desires. Another technique adjusts the conversion factor over a several month period based on howmuchinhabitants are willing to pay for gas andelectricity. Implementation Wehave implemented ACHEin an actual residence. The residence is a former three-roomschool housebuilt in 1905 near Boulder, Colorado, originally serving children of the miningtown of Marshall (Figure 1). The school was closed in 1956 and was completely renovated in 1992, at which time the infrastructure needed for the ACHEproject was incorporated into the house, including nearly five miles of low-voltage conductor for collecting sensor data and a power-line communicationsystem for controlling lighting, fans, and electric outlets. Theresidence is an ideal candidate for intelligent energy managemenbtecause of its age, 13-25 foot ceilings, and exposedsouth and west faces that hold potential for passive solar heating. ACHiEs equipped with sensors that report the state of the environment. The sensory state includes the following for each roomin the home: Ї status of lights (on or off, andif on, intensity level) Ї status of fans (speed) Ї status of temperature control user interface (a fancy digital thermostat that specifies the current setpoint
r+ + + ! ,.....,u,.n.;.I.._._.I.F.l.;.i..... 3
Figure2. Afloorplanof theadaptivheouse,includinglocationsof sensorsandactuators.
temperature for the room, and can be adjusted by the inhabitant) Ї ambient illumination Ї room temperature Ї soundlevel Ї motiondetector activity (motion or no motion) Ї status of all doors and window(sopen or closed). In addition, the systemreceives the following global information: Ї water heater temperature Ї water heater energy usage Ї water heater outflow Ї furnace energy usage Ї outdoor temperature Ї outdoorinsolation (sunlight) Ї gas andelectricity costs Ї time of day, dayof week,date. At present, ACHEhas the ability to control the following actuators: Ї on/off status andintensity of light banks(22 total) Ї on/off status and speedof ceiling fans (6 total) Ї on/off status of water heater Ї on/off status of gas furnace Ї on/off status of electric spaceheaters (2 total) Ї on/off status of speakers in each roomthrough which computer can communicate(12 total) Figure 2 showsa floor plan of the residence, as well as the approximatelocation of selected sensors and actuators.
ACHE Architecture Adaptivecontrol of building energysystemsis 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 mustbe controlled simultaneously. Undersuch circumstances, traditional techniquesfromcontrol theory and Artificial Intelligence have great difficulty (Dean&Wellman,1991). The basic system architecture of ACHiEs presented in Figure 3. This architecture is replicated for each control domain--lighting, air heating, water heating, and ventilation. Theinstantaneous environmentalstate is fed througha state transformation that computesstatistics such as averages, minima, maxima,and variances in a given temporal window.The result is a state representation that provides moreinformation about the environmentthan the instantaneous values. The instantaneous state is also given to an occupancy model that determines for each zone of the house--usually corresponding to a room--whetheror not it is occupied. The occupancymodel relies on motion detector signals, but it includesrules that say, essentially, "a zone remains occupied, even when there is no motion, unless there is motionin an adjacent zone that wasprevious unoccupied." Consequently, the occupancy model maintains occupancystatus even whenthere is no motion. The three adaptive components of ACHEare shownin the top of Figure 3. Variouspredictors attempt to take the current state and forecast future states. Exampleosf predictions include: expected occupancy patterns in the house over the next few hours, expected hot water usage, likeli-
~decision I device regulatoIr Tsetooint prdfile setpoint IgeneratoIr
l information ictors
occupied zones
I sf°tocarctmeutpraaatninoscn"yI I moael I
4L I instantaneous environmenstatal te
Figure3. Systemarchitecture of ACHE
hood that a zone will be entered in the next few seconds. Thepredictors are implementedas feedforward neural networks trained with back propagation, or as a combination of a neural net and a look up table. Giventhe predictions of future states, control decisions need to be madeconcerning the energy devices in the home. Thedecision makingprocess is split into twostages. The setpoint generator determines a setpoint profile specifying the target value of someenvironmental variable (lighting level, air temperature, water temperature, etc.) over a win- dowof time. Thedevice regulator controls physical devices to achieve the setpoint. The device regulator mayhave manyalternative devices at its disposal. It mustdetermine whichone or whichsubset to use. Thereason for dividing control betweenthe setpoint gen- erator and device regulator is to encapsulate knowledge. Thesetpoint generator requires knowledgeabout inhabitant preferences, while the device regulator has knowledge about the physical layout and characteristics of the environmentand controlled devices. If the inhabitants or their preferences changeover time, only the setpoint generator need relearn. The setpoint generator and device regulator in each domainare based on one of twoapproachesto control: indirect control using dynamicprograming and models of the environmentand inhabitant, or direct control using Reinforcement Learning. For example, the device regulator for Indoor Air temperatureuses a predictive modelof the indoor air temperature, as a function of the current indoor temperature, outdoor temperature, and the states of the furnace and electric space heaters. This modelis based on a simple RC thermal modelof the house and furnace, with a neural networkthat learns deviations from this simple modeland the
actual behavior of the house. Giventhis model,achieving a particular setpoint temperature involves little morethan exhaustively searching through the space of heating device actions and finding an action sequencethat achievesthe setpoint. In contrast to this indirect approach,the setpoint generator for the lighting controller uses a direct approachwith reinforcementlearning becauseit wouldbe difficult to learn an explicit modelof inhabitant preferences. CurrentImplementationStatus Wehave conducted simulation studies of the heating control system (Mozer, Vidmar, & Dodier, 1997), using actual occupancydata and outdoor temperature profiles, evaluating various control policies. ACHErobustly outperforms three alternative policies, showinga lower total (discomfort plus energy) cost across a range of values for the relative cost of inhabitant discomfort and the degree of nondeterminismin occupancypatterns. Wehave also implementedand tested a lighting controller in the house(Mozer&Miller, in press). Togive the flavor of its operation, wedescribe a samplescenario of its behavior. Thefirst time that the inhabitant enters a zone (we'll refer to this as a trial), ACHdEecides to leave the light off, based on the initialization assumptionthat the inhabitant has no preference with regard to light settings. If the inhabitant overrides this decision by turning on the light, ACHiEmmediately learns that leaving the light off will incur a higher cost (the discomfort cost) than turning on the light to someintensity (the energycost). Onthe next trial, ACHdEecides to turn on the light, but has no reason to believe that one intensity setting will be preferred over
another. Consequently, the lowest intensity setting is selected. Onany trial in whichthe inhabitant adjusts the light intensity upward, the decision chosen by ACHwE ill incur a discomfortcost, andon the following trial, a higher intensity will be selected. Training thus requires just three or four trials, andexploresthe spaceof decisionsto find the lowest acceptable intensity. ACHEalso attempts to conserve energy by occasionally "testing" the inhabitant, selecting an intensity setting lowerthan the setting believed to be optimal. If the inhabitant does not complain,the cost of the decisionis updatedto reflect this fact, andeventually the lowersetting will be evaluatedas optimal. Evaluating ACHE It is our conviction that intelligent control techniques for complex systems in dynamic environments must be developedand evaluated in naturalistic settings such as the Neural Network House. While there are numerous examples illustrating the potential of neural nets for control of building 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. Weintend to show that adaptive control will yield benefits in natural environmentsunder realistic operating conditions. The research program hinges on a careful evaluation phase. In the long term, the primary empirical question we mustansweris whetherthere are sufficiently robust regularities in the inhabitants' behavior that ACHcEan benefit from them. On first consideration, most people conclude that their daily schedules are not "regular"; they sometimes comehomeat 5 p.m., sometimesat 6 p.m., sometimesnot until 8 p.m. However,even subtle statistical patterns in behavior--suchas the fact that if one is not homeat 3 a.m., one is unlikely to be homeat 4 a.m.--are useful to ACHE. These are patterns that people are not likely to consider whenthey 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 Weare grateful to MarcAndersonand 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 (SHARPo) f Sensory Inc., as well as a CRCWgrant-in-aid from the University of Colorado, NSFaward IRI-9058450, McDonnell-Pewaward 97-18. References Curtiss, P. S., Kreider,J. E, &BrandemuehMl, .J. (1994).Local and global control of commercialbuilding HVAsCystems using artificial neural networks.Proceedingsof the 1994 American Control Conference ACC'94.NewYork: IEEE Press. Dean,T. L., &WellmanM, .R(1991). Planningandcontrol. San Mateo, CA:MorganKaufmann. GregorekT, . (1991).Theenergyrevolution. ElectronicHouse,6, 10-15. Miller, R.C., &SeemJ, . E. (1991).Comparisonf artificial neural networkswith traditional methodsof predicting return time fromnight or weekendsetback. ASHRATEransactions,97. Mozer,M.C., Dodier,R., &Vidmar,S. (1997). TheNeurothermostat: Adaptivceontrolof residential heatingsystems.In M.C. Mozer,M.I. Jordan, &T. Petsche(Eds.), Advanceisn Neural InformationProcessingSystems9 (pp. 953-959).Cambridge, MA:MITPress. MozerM, .C., &Miller, D. J. (in press). Parsingthe stream time: Thevalue of event-basedsegmentationin a complex, real-world control problem.In M.Gori(Ed.), Adaptiveprocessingof temporailnformationS. pringerVerlag. Scott, G. M., Shavlik,J. W.,&Ray,W.H. (1992).RefiningPID controllers usingneural nets. In J. E. MoodyS,. J. Hanson&, R. P. Lippman(nEds.), Advancesin NeuralInformationProcessing Systems 4 (pp 555-562). San Mateo, CA:Morgan Kaufmann. SeemJ, . E., &Braun,J. E. (1991).Adaptivemethodfsor real-time forecasting of building electrical demand.ASHRATEransactions 97.

MC Mozer

File: the-neural-network-house-an-environment-hat-adapts-to-its-inhabitants.pdf
Title: The Neural Network House: An Environment that Adapts to its Inhabitants
Author: MC Mozer
Author: Michael Mozer
Published: Sun Jan 20 23:30:48 2002
Pages: 5
File size: 0.47 Mb

, pages, 0 Mb

, pages, 0 Mb

2018 Ministry Guide, 15 pages, 0.23 Mb

Archive, 27 pages, 0.94 Mb
Copyright © 2018