Mozer, M. C. (1999). IEEE Intelligent Systems, in press.

An intelligent environment must be adaptive

What will the home of the future look like? One popular vision is that household devices -- appliances, entertainment centers, phones, thermostats, lights, etc. -- will be endowed with microprocessors allowing the devices to communicate with one another and with the home's inhabitants. The dishwasher can ask the water heater whether the water temperature is adequate; inhabitants can telephone home and remotely instruct the VCR to record a favorite show; the TV could select news stories of special interest to the inhabitant; the stereo might lower its volume when the phone rings; and the clothes dryer might make an announcement over an intercom system when it has completed its cycle.

The cost of the hardware infrastructure is not prohibitive if the devices are mass produced and if communication is conducted over power lines or wireless channels; and even if the cost is too high today, wait a few years and the price will drop precipitously. Adopting a uniform communication protocol is also not an obstacle in principle. The real reason why this vision of home automation seems unlikely is that it requires a significant programming effort, and worse, the programming must be tailored to a particular home and family and must be updated as the family's lifestyle changes. Tackling the programming task is far beyond the capabilities and interest of typical home inhabitants. People are intimidated by the chore of programming simple devices like VCRs and setback thermostats, never mind a much broader array of devices with far greater functionality.

Perhaps if you were Bill Gates, you might hire a full-time team of engineers to customize your system and keep it up to date. Some commercially available systems adopt this strategy on a smaller scale: Following installation, a technician comes to the home, consults with the inhabitants, and sets up the initial programming. As the inhabitants' needs change over time, the technician can modify the programming, either remotely or on site.

The Adaptive House

In contrast to existing automated homes that can be programmed to perform various functions, the crux of our research is to develop a home that essentially programs itself by observing the lifestyle and desires of the inhabitants and learning to anticipate their needs (Mozer, 1998; Mozer & Miller, 1998; Mozer, Vidmar, & Dodier, 1997). The intelligence of this house lies in its ability to adapt its operation to accommodate the inhabitants; thus, we call the project the adaptive house.

Traditional automated homes require a user interface, such as a touch screen which displays the system state and controls, or a speech recognition front end. However, even a well-designed interface is an impediment to the acceptance of an automated home. In contrast, the adaptive house should be unobtrusive and require no special interactions. Inhabitants operate the adaptive house as they would an ordinary home -- using light switches, thermostats, on/off and volume controls like those to which they are accustomed. Unlike an ordinary home, however, these adjustments are monitored and serve as training signals -- indications to the house as to how it should behave.

The adaptive house infers appropriate rules of operation of devices from the training signals and from sensors that provide information about the environmental state. As the house becomes better trained, it begins to anticipate the inhabitants' needs and set devices accordingly, gradually freeing inhabitants from manual control of the environment. For example, it could automatically maintain the room temperature to a level appropriate given the particular occupants, activities, manner of dress, and time of year; it could choose one pattern of lighting while dinner is being prepared, and another for a late-night snack; it could turn on the television news during dinner, or play classical music when water is drawn for a bath, based on past selections of the inhabitants. Ideally, the house's operation is transparent to the inhabitants, other than the fact that they do not have to worry about managing the various devices in the home.

One might view the home as a type of intelligent agent that infers the inhabitants' desires from their actions and behavior. Intelligent software agents abound that attempt to satisfy the information needs of users of the world-wide web. We extend the idea of the intelligent agents to comfort needs of people in natural living environments. Intelligent agents seldom perform perfectly because of their limited ability to infer users' intentions. However, this problem can be minimized in natural environments through the use of smart sensors and some general domain knowledge, e.g., an analysis of typical tasks performed in an environment.

Residential Comfort Systems

To discuss the adaptive home in concrete terms, let us focus our discussion on the control of basic residential comfort systems: air heating, lighting, ventilation, and water heating. The reason for this focus is twofold. First, these devices are prime consumers of energy resources, and thus present an opportunity for energy conservation. Second, although some devices in the home are controlled with relative ease, such as the stereo or TV, the operation of residential comfort systems can be quite complex. Consider the control problem for air heating. (In Colorado, we worry more about heating in the winter than cooling in the summer.) The thermostat could be set to 70 around the clock. However, this is inefficient because the house does not have to be heated while its inhabitants are at work, nor does the setpoint have to be as high at night. A digital thermostat with multiple setback periods could be used to specify when to lower the setpoint. However, different rules are required for weekdays and weekends. Further, the setback thermostat only allows one to specify first-order rules of occupancy (expected departure/return time based on weekday versus weekend); for efficiency, the thermostat should really know more subtle patterns of occupancy (expected return time based on day of week, departure time that day, weather conditions, recent schedule, etc.). It must also take into account the time required to heat the house, which depends on outdoor weather conditions. If the house has multiple furnaces or zoned control, room occupancy patterns must be taken into account. Further, alternative means of heating should be considered, e.g., opening blinds to allow for passive solar gains, electric space heaters to heat individual rooms, or fans to mix the air. Finally, utilities may charge for energy based on time of use, making it more efficient to overheat the house during the day than to heat it to the appropriate setpoint immediately before the return of the inhabitants. Thus, it is a nontrivial challenge to regulate the air temperature in the house so as to simultaneously maintain comfort and energy efficiency.

ACHE

We have constructed a prototype system in an actual residence. The residence was completely renovated in 1992, at which time the infrastructure needed for the adaptive house project was incorporated into the building, 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.

We call the system that runs the home ACHE, an acronym on Adaptive Control of Home Environments. At present, ACHE has the ability to control 22 banks of lights (each having 16 intensity levels), 6 ceiling fans, 2 electric space heaters, a water heater, and a gas furnace. ACHE is equipped with roughly 75 sensors, which include the following for each room in the home: intensity setting of the lights, status of fans, status of digital thermostat (which is both set by ACHE and can be adjusted by the inhabitant), ambient illumination, room temperature, sound level, status of one or more motion detectors (on or off), and the status of doors and windows (open or closed). In addition, the system receives global information such as the water heater temperature and outflow, outdoor temperature and insolation, energy use of each device, gas and electricity costs, time of day, and day of week. Figure 1 shows a floor plan of the residence, as well as the approximate location of selected sensors and actuators.

Objectives of ACHE

ACHE has two objectives. One is anticipation of inhabitants' needs. Lighting, air temperature, and ventilation should be maintained to the inhabitants' comfort; hot water should be available on demand. If inhabitants manually adjust an environmental setpoint, they are indicating that their needs have not been satisfied. The second objective of ACHE is energy conservation. Lights should be set to the minimum intensity required; hot water should be kept at the minimum temperature needed to satisfy the demand; only rooms that are likely to be occupied in the near future should be heated; when several options exist to heat a room, the one minimizing expected energy consumption should be selected.

Achieving either 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. In what sort of framework can the needs of the inhabitants be balanced against energy conservation? 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 anticipated by ACHE. An energy cost is incurred based on the use of gas and electricity. ACHE's goal is to minimize the combined costs of discomfort and energy.

This framework requires that discomfort and energy costs be expressed in a common currency, which we have chosen to be dollars. Energy costs can readily be characterized in dollars, but some creativity is involved in measuring discomfort costs in dollars. Relative discomfort is indicated when the inhabitant manually adjusts a device (e.g., turning on a light), and this relative discomfort is translated to a dollar amount by means of a misery-to-dollars conversion factor. One technique we have used to specify this factor is based on an economic analysis in which we determine the dollar cost in lost productivity that occurs when ACHE ignores the inhabitants' desires. Another technique adjusts the conversion factor over a several month period based on how much inhabitants are willing to pay for gas and electricity.

Prediction and Control

To minimize combined discomfort and energy costs, ACHE requires the ability to predict inhabitant lifestyle patterns and preferences, and to model the physics of the environment. We illustrate with a simplified scenario: It's 6:00 p.m. and the home is unoccupied. ACHE must decide whether to run the furnace. On the one hand, if the furnace is turned on for the next half hour, ACHE predicts that the indoor air temperature will rise to a level tolerable by the inhabitants, but an energy cost will be incurred. On the other hand, if the furnace is left off and the inhabitants return home around 6:30, a discomfort cost will be incurred. The decision about the furnace state depends on the expected energy cost on the one hand and the expected discomfort cost on the other hand (which depends on the probability that the inhabitants will return).

ACHE thus requires predictors which estimate future states of the environment such as the probability of home occupancy (based on 30 variables, including recent occupancy patterns) and indoor air temperature (based on outdoor air temperature and a thermal model of the house and furnace), as well as inhabitant preferences such as the temperature required to keep the inhabitant from "punishing" ACHE. The predictors are based in large part on neural networks, which are statistical pattern recognition devices inspired by the workings of the brain. Neural networks have the ability to learn from experience -- from data collected by the house.

We have conducted simulation studies of the air heating system (Mozer, Vidmar, & Dodier, 1997), using actual occupancy data and outdoor temperature profiles, evaluating 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 nondeterminism in occupancy patterns.

We have also implemented and tested a lighting controller in the house (Mozer & Miller, 1998). To give the flavor of its operation, we describe a sample scenario of its behavior. The first time that the inhabitant enters a room (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 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 conserve 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. As described in this scenario, ACHE relies on reinforcement-learning techniques (Sutton & Barto, 1998). ACHE includes a neural network that predicts when a room is about to become occupied, so that the lighting can be set prior to room entry. The scenario presented sidesteps a difficult issue: lighting preferences depend on the context (time of day, current activities, ambient light level, etc.), thus requiring ACHE to learn about desired lighting patterns in a context-dependent manner.

Discussion

Our 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 statistical regularities 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, higher-order 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 unquestionably present, and our experiments to date suggest that they can be exploited to serve as the foundation of an intelligent, adaptive environment.

Acknowledgements

We are grateful to Marc Anderson and Robert Dodier, who helped develop the software infrastructure for the Adaptive House. This research has been 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, McDonnell-Pew award 97-18, and NSF awards IRI-9058450 and IBN-9873492.

References

Mozer, M. C., Vidmar, L., & Dodier, R. H. (1997). The Neurothermostat: 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. (1998). Parsing the stream of time: The value of ev ent-based segmentation in a complex, real-world control problem. In C. L. Giles & M. Gori (Eds.), Adaptive processing of temporal sequences and data structures (pp. 370-388). Berlin: Springer Verlag.

Mozer, M. C. (1998). The neural network house: An environment that adapts to its inhabitants. In M. Coen (Ed.), Proceedings of the American Association for Artificial Intelligence Spring Symposium on Intelligent Environments (pp. 110-114). Menlo Park, CA: AAAI Press.

Sutton, R. S., & Barto, A. G. (1998). Reinforcement learning: An introduction. Cambridge, MA: MIT Press.