WO2007149064A1 - Method for tracking using dynamic relational bayesian networks - Google Patents

Abstract

A method for monitoring real-time behavior of a system includes providing a dynamic relational Bayesian network (DRBN)1 developing an instance network for tracking time with respect to dynamic variables, employing a filtering method to draw infererences from the DRBN network, the system applied to the modeling of in-resident patient, moving and physical targets, or automobiles in traffic. The system provides modeling and efficiency improvements over representations described in the past.

Description

METHOD FOR TRACKING USING DYNAMIC RELATIONAL BAYESIAN NETWORKS

CROSS REFERENCE TO RELATED APPLICATION

This application claims the benefit of U.S. Provisional Application No. 60/668,376, filed April 5, 2005.

Description

Elements:

1. A Dynamic Relational Bayesian Network

2. An efficient representation for the instance network

3. An efficient approximate filtering algorithm for large, multi-level, mixed discrete-continuous relational tracking models

4. A method for detecting and terminating instances of dynamic entities and relations (optional)

1. Our DRBN representation is based on our RBN representation, which has been described in the past, and includes the use of our patent-pending synthetic variable language. Three extensions not listed in previous patents include:

1. The ability to designate certain nodes in a DRBN as import, export, or update. Each import node must resolve to an export node in another model, and each instance node referred to from multiple models must be labeled as an export node in exactly one model. For example, if Session. cluster is referenced in a Click (e.g., as Click.session. cluster), then Session. cluster must be labeled as an export node in the session model, and labeled as an import node is any Click models in which it is referenced.

2. the ability to define multiple sub-models for a class (entity or relation), and associate a condition of applicability, expressed as a Boolean-valued synthetic variable, to each submodels. Only those model(s) for which the applicability condition evaluates to true are applied to each instance.

3. Classes are labeled as static, dynamic, or dependent. The use of this labeling will be described in the filtering algorithm description below.

2. In our target applications most variables in most models are known at tracking time. Our tracking representation consists of two elements: a reduced graph and a set of instance node caches.

1. The reduced graph is a copy of the DRBN in which each variable that will be known at tracking time is reduced to a singleton domain. This copy is then compiled to compute all marginals and subjoints (note - this is an extension of the methods of Bloemeke and Valtorta - attach DCMM ppr too!). A final extension is the use of dependency tracking to enable rapid, memory-allocation-free, re-evaluation of the compiled models.

2. A cache is constructed for each instance node in the unrolled graph. This cache includes a. the node posterior, b. the product of any received likelihood messages, c. the likelihood message sent to the node parent (if the node is an import node) the node value (if observed), and d. an array of indicies, 3. Note that cache elements are constructed even for variables labeled as "import" in the class graph - this results in automatically creating a separate likelihood store for each child of the instance node. The array of indicies maps from the full distribution defined in the class graph node to the subdistribution needed for inference given the observed values for this instance node and its parents. These arrays are themselves uniquified to save memory, and permit very rapid replacement of the probability tables in the compiled reduced class graph as it is re-used for each instance.

3. Our filtering algorithm solves both space and time issues in relational tracking, and consists of the following components:

1. Generalized Loopy inference (Ghouramani), using individual DRBN instance models as the unit for loopy inference, and compiled SPI within individual instance models (Bloemeke and Valtorta). We reverse the order of instance application for each loopy pass. This greatly reduces the sensitivity of loopy convergence to instance creation order, (essentially this effects a "forward-backward" algorithm that works well for parameter estimation as well as filtering)

2. Dynamic adjustment of the set of instances over which GLI is applied at each time step - GLI is applied to all instances created in the n most recent time-steps, where n is a parameter of the algorithm.

3. Rollup of instances is accomplished by checking, following loopy inference over new instances, if any new instance: (1) is of a class labeled "dynamic"; and (2) has any nodes labeled "import" AND "update". If this is the case, then the instance containing the export node corresponding to the identified import node is located and the procedure is recursively applied. This procedure is repeated n times, where n is a parameter of the algorithm. The instance at the end of this recursion, if any, is scavenged: that is, all of its node caches, as well as all of the node caches for any nodes that are children of a scavenged node and are owned by instances of classes labeled dynamic or dependent. No special inferential computation need be done since remaining nodes, if they are children of scavenged nodes, already have posterior marginals (a property of GLI), and if parents, already have the likelihood from the scavenged child incorporated into the parent likelihood.

4. Finally, our dynamic instance detection and termination method relies on the use of user-specified policies. These policies can be specified using arbitrary Java code or as synthetic variables. These synthetic variables are attached to database insertion events, and can test arbitrary conditions including probabilities of unobserved values for any instance that can be referenced from the triggering instance.

A. Introduction

Technological Innovation

The technological innovation of the proposed work is the application of dynamic relational Bayesian networks (DRBNs) to activities in the home environment. DRBNs are an instance of a new family of approaches emerging in the machine learning community based on a relational data model. DRBNs are well-suited for modeling in-home activity because they can directly model relationships among people, places, objects, activities, and sensed phenomena. DRBN algorithms exploit the data model and meta-data from the schema to guide and frame relational queries about behavior and events. These provide the basis for description of complex events (such as awakening in the morning) and also provide a way to formalize status assessments. The DRBNs formulated can represent complex, dynamic, multi-scale processes involving multiple actors, as probability distributions over the elements, queries, and relationships in the DRBN model.

DRBNs also provide the means to balance the concerns of privacy and sensitivity. Privacy concerns are addressed by the local processing of sensor data. Only Events Requiring Intervention (ERIs) will be forwarded on the caregiver or family member for action. Sensitivity concerns are addressed by algorithms that fuse sensor data with risk assessment to formulate a realistic assessment of the significance for the event under consideration.

Commercial Application

The aging of the U.S. population presents many challenges. The existing paradigm of care will not allocate resources efficiently as the size of the population requiring home care assistance grows. Technology currently commonplace for security monitoring and remote household management may provide tools that will extend home health care resources by monitoring elders in their homes and inferring occurrence of events and behaviors requiring intervention. Our goal is to enhance elder independence by providing both better, more timely, predictive, health-status assessments and direct, real-time, recommendations and warnings. These together will lower the risk of elders remaining at home or in low-intensity care settings.

Proposed Product

The proposed product, CleverSet Activity Tracker (CAT), will deliver activity summaries, recommendations, and risk assessment reports. We plan a staged roll-out of products to two market segments, institutional and home-care managers and the elder's family. The product for both market segments includes a drop-in set of sensors and processing of the sensor stream. CleverSet will outsource sensor installation. For large institutional buyers, CleverSet will place a server system on-site and provide maintenance services under contract. For home-based customers, CleverSet will outsource installation of a low-cost home computer and sensors and license the software to provide recommendations, and risk assessment reports.

The result of the Phase Il research will be a CAT prototype that incorporates functional algorithms, interface, and analyses, and that demonstrates the functionality of the technology targeted to aging in place.

B. Significance and Related R&D

For many elderly individuals, technology is both a blessing and a curse. In medical care, it can be a confusing and intrusive reminder of one's growing dependence on others. It also can be a harbinger of future interactions with the technology-dense hospital environment, where many elders receive end-of-life care.

Technology also is a lifeline to the outside world. Email and electronic media provide an economical and timely means for elders to communicate with friends and family and to maintain an awareness of current events. It also provides a means to summon help when it is needed.

In this Phase Il proposal, we present technology that is unobtrusive, actionable and respectful of privacy. This work will allow elders to live independently longer, in their own homes, in continuing-care retirement communities, and assisted living facilities. The work differs from other industrial and academic work in that it is broad-based. Rather than testing the limits of hardware (i.e., sensors and monitoring equipment), it uses a set of simple, robKt»sens6rfe^|ia'l^s(tl¹«ptoit#tl^tte'aata^»<5i3taineci from them computationally. The goal is to discern the activities of daily living, to detect events requiring intervention, and to report them

We now review existing knowledge in the domains of ADL assessment, dynamic relational Bayesian networks, relational influence diagrams, sensor models of activities of daily living and targeted risk assessment. Table B- 1 shows the applicability of the related work reviewed to the Specific Aims of the proposal.

Table B-1. Related work and relevance

B.1 Activities of Daily Living (ADLs)

Activities of daily living (ADLs) are the common, everyday tasks associated with independent living The most common instrument used to measure a person's functional ability to accomplish these tasks is the Katz Activities of Daily Living Scale (Katz, 1963). This scale assesses bathing, dressing, transferring, using the toilet, continence, and eating. These ADLs, however, do not measure the full range of activities necessary for independent living. Instrumental activities of daily living (IADLs) were developed to capture a more complex range of activities, including handling personal finances, meal preparation, shopping, traveling, doing housework, using the telephone, and taking medications (Lawton and Brody, 1969). IADLs may be difficult to use in an institutional setting, where tasks such as meal preparation and housework are done by others. In a home setting, an individual's inability to perform an IADL, meal preparation, for example, may be due to the fact that the spouse performs this activity. ADL and IADL instruments have been freely modified over the years (DHHS 1989). In this spirit, we propose to model a mixed set of ADLs and IADLs in the proposed work.

ADL measurements have been found to be significant predictors of nursing home admissions (Hanley et a/, 1989), home care services usage (Newman et a/, 1988), hospital care usage (Liu and Manton, 1988), living arrangements (Bishop 1986), Medicare expenditures (Gruenberg and Tompkins, 1986), insurance coverage (Dunlop, et a\, 1989) and mortality (Manton, 1988). ADL measurements are now a standard variable included in gerontological research. ADL measurements are also important because an increasing number of private long-term care insurance policies and proposed public long-term care insurance programs rely on ADL assessment to determine whether or not an individual qualifies for benefits. John Hancock, Aetna, Travelers, Metropolitan Life and CAN sell private insurance policies that rely on ADL measurements to trigger benefits (Van Gelder, in DHHS 1989).

In the SBIR Phase I effort CleverSet developed a methodology to track ADLs from sensor data using DRBNs. These models are superior to current ADL assessments because (1) they provide time intervals for beginning and end of the ADL; (2) when expanded to a more complete set of ADLs, they provide objective coverage throughout the day, not based on the elder's subjective recollection or a short interview/home visit by the caregiver, and (3) they form a record that can be examined and referenced by the individual monitored, his or her family, and caregivers.

B.2 Dynamic relational Bayesian networks (DRBNs)

Sensor reports are noisy and unreliable. Motion sensors have uneven and possibly overlapping regions of coverage, and are prone to false alarms (e.g., being triggered by a wind-blown window curtain) and missed detections (e.g., below-threshold motions, low batteries, or communication failures). Light sensors can be triggered by changing shadows from sunlight. The standard solution to these problems is sensor fusion, both over sensors and over time. Monitoring of objects/events over time based on sensor input is termed tracking. Bayesian networks (BNs) provide a compact way to describe the joint probability distribution over situation variables (Pearl, 1988), and so provide a natural and increasingly popular basis for sensor fusion. Dynamic Bayesian netwotfcs (BoyefV€nd'Kbllerf 1^i;999)^>>exfend BNs into dynamic domains and have been widely adopted for modeling and interpreting time-varying data in speech, robotics, target-tracking, and other domains (e.g., Crisan and Doucet, 2002). With Takikawa (Takikawa et al, 2002), we have studied exact methods for tracking using large-scale DBNs with simple hierarchical structure. Marthi (Marthi et al, 2002) has applied MCMC techniques to filtering (tracking) in DBNs. Particle filtering (Doucet et al, 2000) is a popular approach for low- dimensional problems such as location/orientation tracking.

The above approach is adequate when there is a single level of entity to be tracked (e.g., tracking an individual within a household), but does not readily scale to more comprehensive, multi-scale situation descriptions. For example, imagine a motion sensor report from the apartment common area at 2am. Simple tracking is sufficient to estimate that the occupant may now be in the common area. A more complete description is multilevel:

1. The occupant is in the common area

2. The occupant has moved from the bedroom to the common area

3. The occupant is about to use the toilet

4. The occupant is in a sleep-interruption-for-toileting.

This last is the ADL-relevant level of situation description, and is not obtainable through traditional sensor- fusion.

Relational Bayesian Networks (RBNs) extend Bayesian network technology to a relational data model, permitting modeling of situations where relationships are important (e.g., between people and locations) and/or where varying numbers of entities (people, activities) can be occurring (Heckerman, 2004; Getoor, 2001; Koller, 1999; Friedman ei a/., 1999; Neville and Jensen, 2000). Recent research Dynamic Relational Bayesian Networks (DRBNs) extends the applicability of RBNs to dynamic (time-varying) problems (D'Ambrosio, 2005; Milch et al., 2004; Sanghai et al, 2003 ).

CleverSet is a world-expert in the application of DRBNs to behavioral monitoring from digital data (D'Ambrosio et al, 2003). CleverSet has applied proprietary DRBN technology in a number of military and scientific domains (D'Ambrosio, 2004), and will leverage this knowledge and experience in the Phase Il effort.

B.3 Sensor models of the activities of daily living

B.ZΛ MIT

In collaboration with the Changing Places/House_n Consortium, researchers developed a system for recognizing activities in the home setting using a set of small and simple state-change sensors. The system was designed to be quickly and ubiquitously installed in home environments, consisting of a set of simple, inexpensive sensors that can be affixed to objects with physically manipulated open-close or on-off states, such as doors, light switches, appliances, cabinetry, and containers. These sensors record timestamps in response to actions inhabitants take on the environment, such as opening the fridge. Drawback: These sensors record data internally for the duration of the experiment, after which the sensors must be collected for the data to be uploaded, synchronized and analyzed Therefore, this system provides no means by which to deliver data for real-time situation assessment or action. The system has been deployed in multiple residential environments with non-researcher occupants. Preliminary results on a small dataset show that it is possible to recognize activities of interest to medical professionals such as toileting, bathing, and grooming with detection accuracies ranging from 25% to 89% depending on the evaluation criteria used (Tapia and Intille, 2004).

B.3.2 Oregon Health and Sciences University

Preliminary research in the assessment of motor activity and walking speed using a system based on wireless motion sensor technology. Initial data from Oatfield Estates, an assisted living facility run by Elite Care, Inc., which monitors occupants with variety of sensors to record locations of residents and staff, motion in residents' apartments, weight on residents' beds, temperature, etc. Drawback: this facility was designed and built using robust industrial sensing and structured wiring technologies, but without the benefit of behavior modeling nethodologiesVwhidMitfflsfts^tility a"s"a' research test-bed for iterative design. These data are used in the ievelopment of inference models. Current projects include describing the stochastic relationship between actions and sensor observations in the form of a Hidden Markov Mode (HMM), and assessing mobility based jn analysis of the patterns of movement between location sensors (Hayes et al, 2003). As we see in the Phase report, HMMs do not scale to the broader task of detecting and assessing ADLs.

3.3.3 Intel n Philippose ef al (2003), Radio Frequency Identification (RFlD) sensors are attached to the person and many objects in the environment. Recognition of activities is accomplished via an RFlD reader-glove when tagged objects are manipulated. Drawbacks: (1) all items must be tagged; (2) requires the user to wear a reader-glove to interrogate the RFID tags.

B.3.4 Honeywell International

The Independent Lifestyle Assistant™ (I.L.S.A.) is an agent-based monitoring and support system to help elderly peopie to live longer in their homes by reducing caregiver burden, I.L.S.A. is a multi-agent system that incorporates a unified sensing model, situation assessments, response planning, real-time responses and machine learning. A six-month study of the system, deployed in elder's homes, included continuous data collection and transmission via security sensors installed in the home, data analysis, information synthesis, and information delivery to I.L.S.A. clients and their caregivers. The test concentrated on monitoring two of the most significant Activities of Daily Life (ADLs): medication and mobility. Sensor types: each of these installations had between 10 and 20 sensors, including motion sensors in every room, door contact switches on all exit doors, medication caddies, pressure mats in 'strategic' locations such as the bathroom sink, flush sensors, a medication caddy, and security sensors. Significant concluding quote: "Technology will not replace hands-on human caregiving, but it will allow us to direct those scarce and expensive human resources more effecf/Ve/y" (Haigh ef al, 2004).

Another issue in formulating models of activities of daily living is the origin of the schema. Several approaches have been explored, including a pre-formed ontology (Haigh et al, 2004). The Consolidated Home Ontology in Protege (CHOP), an ontology containing over 800 distinct concepts, was assembled for use in I.L.S.A. CHOP was derived from two upper ontologies, Cyc and SUMO (Suggested Upper Merged Ontology). It was developed with Protege, a popular visual ontology construction tool, and through mining the world wide web for common sense knowledge (Perkowitz et al, 2004).

B.3.5 Living lndependently.com

QuietCare is a commercial in-home monitoring and data-reporting system based upon information obtained by five sensors disposed about the user's home. The system detects movement, toileting, and medication-taking data for the elderly, using motion sensors. The system uploads data every two hours to a server, which compares current values with a running baseline. Reports are published to a secure website. Drawbacks: (1) the system cannot independently or immediately assess a particular situation being encountered by the user, nor can it determine occupancy status, upon which all subsequent inferences rely; (2) the system cannot independently determine appropriate actions based upon the sensor data - a caregiver must be trained to analyze the information and perform an appropriate response (Glascock and Kutzik, 2000).

CareWheels has implemented a robust sensor testbed at the Pine Point Apartments in Portland, OR (see Section D.3.2). These data provide a basis for tracking ADLs using continuous, long-term data of real individuals living their daily lives. _^

B.4 Targeted Risk Assessment

One of the more challenging aspects of injury epidemiology is determining the time period of risk in order to estimate the actual risk parameter. Little progress has been made in this area since Haddon, Suchman, and Kline's (editors) landmark 1964 publication, Accident Research Methods and Approaches. For this reason, people are assumed to be at risk at all times for each type of injury regardless of their actual exposure to a potentially hazardous situation. As an example, when estimating rates of kitchen-related falls, expressed as the number of such falls divided by the number of person-years during which the falls occurred, the denominator oHfte rfete^' &3stftM§V2#rϊouf^' p§r~βay period of risk rather than the time during which a person is actually in the kitchen. This approach produces conservative estimates of rates with the degree of the sonservativeness varying from person to person and from one type of fall or other injury to another (Rossignol, 1996). An improved approach would measure, for example, time actually spent in trie functional kitchen triangle (sink to stove to refrigerator and back) for kitchen-related falls.

More than one-third of adults ages 65 years and older fall each year (Hornbrook, 1994; Hausdorff, 2001). Through careful studies, researchers have identified a number of modifiable risk factors:

• Lower body weakness (Graafmans, 1996) t Problems with walking and balance (Graafmans, 1996; AGS 2001)

• Taking four or more medications or any psychoactive medications (Tinetti, 1989; Ray, 1990; Lord, 1993; Gumming, 1998).

These factors represent the types of risk modifiers that will enhance the sensor-based models.

Client-specific risk assessments can be performed in our approach by recognizing high-risk activities, their frequency and duration. For instance, in the example cited previously, a specific "bathroom-related fall risk" could be estimated by measuring the time a person actually spent in and transiting to the bathroom, instead of assuming a 24-hour per day risk interval, and coupling this time with the known frequency of bathroom-related falls.

Current injury statistics are blunt instruments. They are neither creative nor insightful. Using DRBNs, activities performed less frequently but which are associated with a large number of injuries could be identified and targeted for intervention.

B.5 Identify commercial opportunity and societal benefits

Eric Dishman, Manager of Intel's Proactive Health Strategic Research Project (SRP) and Chair of the Applications Health subcommittee of the Intel Research Council, summarized the capabilities and promise of this new generation of sensor data fusion and behavioral modeling: "A lot of these people, with a little bit of help from a system, might be able to keep getting dressed by themselves, might be able to keep fixing a meal by themselves, might be able to keep making tea by themselves and enhance what abilities they have let and try to keep them in their own homes as long as possible. The societal benefit will be improved quality of life and prolonged independence."

The commercial opportunity is substantial. In its Market Survey of Assisted Living Costs (October 2003), MetLife estimated the size of the assisted care market at $336 million growing at 15% per year. B.6 Identify roadblocks that proposed research will overcome

The proposed research will remove the following roadblocks:

• Non-actionable, post-hoc information. Current applications are limited in their ability to provide actionable information in useful time frames. The proposed Phase Il project will enable personalized, appropriate, and timely prompts and assessments.

• Lack of observable data for tracking and inferring health care transitions. Current applications do not track long-term changes that are indicative of changes in health status. The proposed Phase Il project will track general trajectories of change and threshold values that trigger evaluation of health care status.

• Stove-piped products that do not support each other. Currently, products do not have the capacity to interact and are not designed to do so in the future. This is a roadblock because it does not provide an organic solution to health care that can extend to a broader range of serves and health care providers as the technology progresses. The proposed research will provide a foundation for extendable models of care that can easily and efficiently incorporate data from and provide data to other relational data sources.

B.7 Why this research is important? The overarching" feslafcWgBsSfWWerfdbVetdefe to remain in home through intelligent monitoring of behavior. This research will dramatically advance the capabilities of sensor monitoring for home-based care. Currently aroducts exist that notify caregivers and family members when activities have or have not been accomplished _kQuietCare (www.quietcaresystems.com), ePill (www.epill.com)). These work well in settings where groups of aeople are monitored as a group, such as in an institutional setting when a health care manager would like to ι/erify that all occupants of an assisted care facility have awakened, and as a post hoc summary of behavior, such as reassurance for a remote family member that the elder has accomplished certain activities for a given day. However, they fall short of what is needed to enhance elder independence. First, they are not capable of recognizing and recording more complex aspects of simple ADLs, such as periods of restless sleep or nighttime bathroom activity, nor are they capable of recognizing more complex ADLs such as meal preparation. Second, they do not deliver prompts that are integrated with an elder's ongoing daily activities. For example, for the medication task, rather than a reminder activated at a set time every day, a personalized prompt should consider the elder's pattern of activity, for example to deliver a reminder when the elder was near his or her medication station within a reasonable interval of time of meal preparation. Assume that the medication station is in the bathroom and that the elder on average visits the bathroom two or three times in the morning. A predictive system would recognize a bathroom visit 20 minutes prior to a scheduled medication dose, deduce that this is an appropriate interval in which to take this medication, and deliver the prompt when the elder was in the bathroom. That is, the next generation of sensor products should be able to predict behavior and "Every 7 seconds, another baby boomer turns adapt to an individual's rhythms of daily living. Finally, the 50 years old. As they have done in other facets current generation of sensor products also does not monitor of American life, the 75 million people born sensor data to detect long-terms changes in activity. between 1946 and 1964 are about to Reactive care transitions (e.g., following a fall) are often permanently change society again. The sheer over-reactions, and are themselves potentially traumatic and number of people that will be living longer and sudden. Early detection of increasing risk will permit more be more active than in previous generations proactive, rational, and less traumatic care transitions. will alter the face of aging forever. One of the greatest challenges in the new century will be how families, business, and government will respond to the needs, preferences, and lifestyles of the growing number of older adults."

- Joseph Coughliπ, Director, MIT Age Lab

C. Phase I FiffaϊtiRφόrl..

C.1 State the beginning and ending dates for the period covered by the SBIR Phase I grant.

Beginning date: 15 August 2004 Ending date: 31 January 2005

C.2 List all key personnel who have worked on the project during that period, their titles, dates of service, and number of hours devoted to the project.

Table C-1. Phase I Key Personnel

C.3 Summarize the specific aims of the Phase I grant.

The research objective of the proposed work was to fuse in-home sensor data using relational Bayesian networks to infer behavior of occupants in a home setting, that is, to assess Activities of Daily Living (ADLs), and to detect events requiring intervention (e.g., falls). The sensor data were collected by CareWheels Corporation, a Section 501 (c)(3) non-profit public benefit corporation. The CareWheels research project, "Internet-enabled Assistive Technologies for Independent Living and Aging-in-Place" has implemented a home sensor testbed currently operating at an independent living facility for people with severe physical disabilities in Portland Oregon.

The Phase I specific aims were to:

(1) develop a reasonable set of profiles that described suites of sensor outputs associated with types of behaviors. VVe anticipated that these profiles would reflect the temporality of behaviors (getting up in the morning, daytime activities, getting ready for bed), space utilization behaviors (living room behaviors, kitchen behaviors), or resident specific behaviors (wheelchair vs. walker behaviors). These models were be vetted by the CareWheels expert. This was intended primarily as a data-exploration / familiarization step.

(2) develop relational Bayesian network models that could infer behavior (ADLs and events requiring intervention) by modeling the joint probability distribution over the sensors, objects, and actions in the home.

(3) assess the performance of the models using qualitative and quantitative analyses to evaluate the reasonableness of the models based on expert knowledge and the ability of the models to quantitatively infer behaviors from sensor data.

C.4 Materials and Methods (Phase 1)

C.4.1 Data

The data were provided by CareWheels in the form of log files produced by CareWheels sensor system, The files were flat, delimited text files, with no relational structure. The files ranged from a week to four weeks of data with up to 130,000 sensor reports per file.

To convert the data into relational format, an ad hoc PERL program was used to parse the flat text files and reform them into xml elements appropriate to the schema being used for analysis.

C.4.2 Schema construction A series of increasingly ό€mpMnln%ive>^§oh^'effia^Λ»were constructed during Phase I. For the preliminary familiarization phase the schema identified reports, sensors and rooms as the relevant objects of discourse. Activities and events were not specifically named in the schema. Rather, they were modeled as probabilistic spatio-temporal clusters of reports, as will be described later.

The final schema we used for DBRN-based ADL and ERl detection is shown in Figure C- 1. A few notes about this schema:

• Activity refers to the lowest level of activity description, and can be of type sleeping, active, or stationary.

• Location refers to the presence of the client in a room, and so has a start and stop time.

• Move models a change of location., i.e., movement between rooms.

Figure C-1. Expanded schema for ADLs and events requiring intervention detection

Finally, as part of the Phase I activity we began development of a more complete ontology (set of terms and relations) that might serve as the basis for developing a complete schema for ADL and ERI detection and assessment. The ontology, developed in part from study of an earlier in-home monitoring ontology developed by Honeywell, the Consolidated Home Ontology in Protege (CHOP) discussed in Section B.3.4 (Haigh e^ a/, 2004).

' The first step in construction of a relational model is the specification of a Schema, the set of objects in the domain, their attributes, and the possible relations among_them. C.4.3 Modeling environment

Models were created and evaluated using CleverSet Modeler™. Modeler is a robust Java-based research platform capable of constructing and evaluating dynamic and static RBNs. Modeler is capable of expressing arbitrary variables in relational data using a patent-pending synthetic variable language. Modeler is also capable of discovering RBNs from data.

C.4.4 Model formulation

Several levels of models were constructed in the course of Phase I.

• Naϊve Bayes. Naϊve Bayes models are typically used in classification applications. They are Bayesian networks consisting of a single hidden variable representing an unobserved underlying state that captures correlations across sensor observations as "clusters," or values of the hidden variable.

• HMM Naϊve Bayes models, as formulated above, model the situation at a single point in time. Dynamicism was introduced into the models by extending the naϊve Bayes model into a HMM with a single hidden variable. Hidden Markov models are dynamic probability models in which observables (sensor reports) are presumed to be explained, at least partially, by a "hidden" or unobserved state of the system (e.g., whether the client is moving about or stationary), which take into account only the previous state of the system when processing the current time frame.

• DRBN. Hidden Markov models represent the situation as a static, finite, and pre-enumerated set of possibilities whose actual occurrence may vary over time. This works well for small problems, but doesn't extend readily to complex problems where the number of possible situations is not easily enumerable in advance. Dynamic Relational Bayesian Networks (DRBNs) permit the modeling a situation as composed of any number of instances of defined entities (e.g., people, rooms, wakeup events) and relations (e.g, residence by a person in a room, movement between rooms). Our initial thesis was that this level of modeling would be needed to effectively detect and assess ADLs/ERIs

Naϊve Baves Model

Naive Bayes models are static probability models in which observables (sensor reports) are presumed to be explained, at least partially, by a "hidden" or unobserved state of the system (e.g., whether the client is moving about or stationary). As part of our initial exploration we constructed a number of naϊve Bayes models and used unsupervised learning (clustering) with these models to detect commonly occurring patterns in the sensor reports over the apartment. An example learned model is shown in Figure C-2:

• Cluster 0 seems to indicate that there is no activity present in the living room and only slight activity in the hallway The resident is most likely not in the living room.

• Cluster 1 seems to indicate that there is a great deal of activity in the living room that is newer than the activity in the hallway. It is very likely that the resident is in the living room and is active.

Cluster Bed BedM BathM Hall LRM KitchM

IπBed On On On On On

0 50.6 84.2 52 0 2.8 9 2 23.9

1 8.9 17.6 60 6 23 8 426 75.9

2 52.3 39.3 64.9 8.8 9.6 27.9

3 3.2 8.0 61.5 21 8 52 0 62.9

Probability of sensor on by cluster

Figure C-2. Learned naϊve Bayes model of commonly occurring patterns in the sensor reports

Hidden Markov Model

Hidden Markov models are dynamic probability models in which observables (sensor reports) are presumed to be explained, at least partially, by a "hidden" or unobserved state of the system (e.g., whether the client is noving about bf teftohaVypwIifch taKWhtό a'ccbunt only the previous state of the system when processing the current time frame The models were further developed to include additional hidden variables representing activity states in specific locations around the apartment, such as activity in the bathroom. Cluster profiles consisted of an associated series of conditional probability distribution for each observed variable given the selected value of the underlying cluster variable.

In Figure C-3, previous hidden variable states (white) are connected to the current states (light gray). Information used in estimating state distributions on these hidden variables is contained in the observed variables, which are grey. Also, the light gray variables that are not connected to the network represent initial states of the system

The hidden state variable of the HMM model captures specific activities in specific rooms. The plot shown in Figure C-3 represents the changing probability distributions of the hidden variable for bathroom residence over time. That is, each vertical slice represents the probability that the client is in the bathroom, hallway, just entering the bathroom, just exiting the bathroom, or not present.

The domains of the hidden variables represent the target activities we are trying to capture with the model.

Figure C-3. HWlIW (top) and tracks of probability of hidden variable states (by location - bottom)

Dynamic Relational Bayesian Model

Dynamic Relational Bayesian Networks (DRBNs) permit the modeling a situation as composed of any number of instances of defined entities (e.g., people, rooms, wakeup events) and relations (e g, residence by a person in a room, movement between rooms). DRBNs fully exploit the relational nature of the data and schema in an RBN. Additional schema classes were created to represent events of interest, such as wake-up. These new schema elements are fundamentally different from those in previous models. While there is only one bathroom in the apartment, there can be any number of wakeup or fall events. In building our DRBNs, we adopt a racking modell'irf wPiich W&iuøfMerneM'thi schema with track initiation and track termination policies: specific iolicies about when we will introduce a hypothesis that a specific event may be starting and/or terminating. These policies, called triggers are defined in terms of the occurrence of suites of specific sensor reports in the context of class of existing situations. Triggers were created with Java code and can be arbitrarily complex and ;an be probabilistic or deterministic (Figure C-4).

The trigger condition for a fall hypothesis, for example, is continuous occupancy of a room for one standard deviation above average occupancy time for that room. One a fall hypothesis is created, the current DRBN ^"uses information about continued room occupancy, location certainty, and subsequent movement to assess the probability of a fall.

Figure C-4. DRBN with triggers. (The model instantiates hypotheses as conditions are met in the sensor data stream.)

C.4.5 Results and Discussion (Phase 1)

In this section we discuss results related to each of our three objectives (initial profiling of sensor data, inference of ADLs and ERIs, and assessment of model performance) in turn.

Profiles of sensor outputs associated with types of behaviors

Profiles were derived using naϊve Bayes models. We began here because we hoped that a simple model would be expressive enough to discover some level detail about activity. We found these naϊve Bayes models to be adequate at capturing broad swaths of activity which were not necessarily localized in time, but due to the structure of the sensor data were localized in space. For example, a hidden state representing inactivity in the apartment was correlated with both nighttime sleeping activity as well as absence.

We found these naϊve Bayes models to be too crude to effectively infer ADL-level activities. They were unable to distinguish nighttime sleeping activity from when no one was home. Frequent activities tended to be represented by several hidden states precluding clear differentiation of activities

Inferences about behaviors

Since the data were inherently dynamic, it seemed sensible to use dynamic models instead of static naϊve Bayes models. HMMs were chosen because they are easily interpretable and a standard model, the first line of dynamic model in a problem such as this The advantages of the HMM are that they capture activity through time. Furthermore, by manually engineering the models, we can define specific activities as states of the hidden state variable. This, however, also proved to be a disadvantage: reasonable results were obtained only after extensive and time-consuming tuning of probability distributions of the hidden variables. One problem with the HMM approach is that it applies at the raw sensor report time-scale, a scale inappropriate for iescribing brol 1S¹¹AOa. As a result, we were only able to effectively model the lowest evel of activity.

3RBN models differ from HMMs in several important ways. First, target activities are represented not as πutually exclusive states of a hidden variable, but as separate data types. This means that we can consider several possible activities simultaneously. This is accomplished by triggering the generation of a new instance tf an activity and subsequently determining its plausibility using the DRBN. Since the activities are a data ype, they can have various attributes which can be incorporated into the Bayesian network. For example, /vake-up activity may have attributes such location (e.g., from a nap in the living room), start time and end time <\ second advantage of DRBNs over HMMs is that the former also allows for easier collection of statistics about activities. This is because each activity has its own associated set of detected instances stored in a relational database. Standard database queries can be used to evaluate, for example, number of hours of sleep per day.

Figure C-5 shows an example of a set of location, activity, and wakeup hypotheses generated over the course of an 18 hour test file. The thick lines denote more confidence in the generated hypotheses than the thin lines. The test file starts at midnight with the person asleep in the bedroom. About a fifth of the way through the night, activity at the front door generates a new hypothesis that the person is awake and at the front door. Notice, though that the sleeping hypothesis and the bedroom location hypothesis are still present, with slightly diminished likelihoods (our models capture the low likelihood the occupant could have moved from the bedroom to the front door without being detected by the common-area or kitchen sensors). The hypothesis about activity at the front door hypothesis continues at low probability for the remainder of the night, until it is terminated when activity in the bedroom following wakeup obviates the front door activity. The DRBN model is capable not only of creating and maintaining simultaneous hypotheses over time, but also or reinstantiating a hypothesis that has been terminated.

Figure C-5. Location, activity, and wakeup hypotheses generated over the course of an 18 hour test file

The wakeup model consisted of two elements: a hypothesis creation policy and a dynamic probabilistic wakeup model. The hypothesis creation policy instantiates a wakeup hypothesis whenever a sleep activity ends. The wakeup model then looks for a suitable period of sleep prior to the wakeup as well as an exit from the bedroom/bathroom area. Belief in the wakeup activity is also conditioned on the belief in the underlying sleep and localftiKferitiJ

The system detected several late afternoon wakeups. This surprised us until we examined the sensor log: it is apparent from the log that the client often took a nap in the late afternoon.

The computational load of the system is quite modest. In general, we were able to process a full day of sensor ogs for one apartment in 2-4 minutes on a 2.8 GHz Pentium processor.

Performance of models

Performance evaluation consisted of two parts: qualitative evaluation by the domain expert, and quantitative comparison of results against ground-truth and a competitive system, Independent Living's QuietCare.

In the qualitative evaluation, the domain expert was asked to evaluate whether or not the results of the naive Bayes, HMM and DRBN models RBNs produced were consistent with activities contained in the sensor data. In all cases, the results were deemed to be reasonable and reflective of the expert's perception of these situations as an active witness and as interpreted through the sensor logs.

Figure C-6 shows a comparison of the time differences of the estimated wake-up times from ground truth for CleverSet's DRBN model (left in both plots) and QuietCare (right in both plots)². Ground truth was provided by the domain expert based on his interpretation of the sensor logs.

,

First wake-up after 4:00 am Closest wake-up to ground truth

Figure C-6. Comparison of CleverSet and QuietCare results for the ADL, wake-up.

(Left - Deviation from ground truth of wake-up time for wake up times for CleverSet and QuietCare for the first wake-up hypothesis after 4am. Right - potential performance of CleverSet's tracking model when it is allowed to consider the closest hypothesis to ground truth.)

Figure C-S is a box-plot in which the box bounded by the values of the first and third quartiles with the median shown by the line inside the box. The height of the box represents the variation contained in the differences. A narrow box shows a higher degree of confidence in the estimate and corresponds with high performance. The larger the vertical width of the box, the worse the performance of the model.

² For this phase we used a one-month sample dataset, drawn from a period of time during which both the Carewheels and QuietCare sensor suites were in the apartment. ^rigure C-6 (leffflhows^'

fforfr gYd'ϋrTα¹ truth of wake-up time for wake up times for CleverSet and UuietCare for the first wake-up hypothesis after 4am. QuietCare reports one wake-up time for the day from sensor data. CleverSet's tracking model can create multiple wake-up times corresponding with the initiation of nany wake-up hypotheses in a day (these are labeled as sleep interruptions, i.e., nighttime bathroom trips, or rue morning wakeups). Each of the hypotheses has an associated probability of existence which can be used .0 express confidence that a particular hypothesis is true.

CleverSet more accurately captures true wake-up time than does QuietCare. (QuietCare consistently estimates a later wake-up time than indicated by ground truth). Figure C-6 (right) shows the potential performance of CleverSet's tracking model when it is allowed to consider the closest hypothesis to ground truth.

Events Requiring Intervention (ERIs): Detecting a fall

Falls are rare events, and none occurred in our datasetε. To evaluate DRBN ability in fall detection, our sensor expert manually modified a one-month data-set by inserting four simulated falls. Table C-2 summarizes performance.

Table C-2. Events Requiring Intervention (ERI) performance metrics

A total of 13S fall hypotheses were triggered over the one-month test period, an average of 4.5 hypotheses per day. Of these, the vast majority had posterior probabilities below 0.05. Of those with posterior probabilities above 0.1 (all had probabilities > 0.7), four were true falls and one was a false alarm. The false alarm corresponded to a period of time during which the bathroom was occupied continuously for over 1.5 hours, a highly unusual event. We believe this event could have been filtered out by more sophisticated models that, for example, fused activity information (the most likely activity throughout the period was "active") as well as time- of-day conditioned room occupancy expectations.

The remaining eleven events with probability greater than 0.1 were unexpected: our fall model correctly identified periods of non-occupancy. Since these are easily distinguished from true falls (the location is "front door outside" rather than an interior room), we view these as correct detections of non-occupancy rather than fall false alarms. Occupancy is a well-known "hard problem" in residence sensor monitoring. While we didn't evaluate the missed-detection probability for non-occupancy³, we did verify that all detected periods of non- occupancy were detections of true events.

************************_*p»iQ QC PROTECTED SBIR ΠΔTΔ*******************************

This demonstration of feasibility supports three core conclusions:

. (1) Relational Bayesian networks (RBNs) are practical models of activity. RBNs consist of three central elements: data, schema, and probabilistic model. These three components support one another. The schema describes the relationships among data as they occur in real life. The probabilistic model is learned from the data, guided by the schema. RBNs are a concrete characterization of activities as they occur. In addition, RBNs support composition of higher-level ADLs (e.g., wakeup) from lower level activities (e.g., sleep, motion, etc.)

³ Evaluation of the missed detection, or false-negative, rate would have required manually reviewing the entire one-month sensor data- set yet again to identify all true periods of non-occupancy. (2) RBNs p'iFbvTde' l&gWεtFWrtSaiδlέ σf'actfvities^Jbf daily living. In general terms, RBNs aggregated data in ways that were consistent with general patterns of behavior with respect to time and place. On a more detailed level, using RBNs, we were able to track one particular activity, awakening, and one ERI, fall.

(3) RBNs provide an accurate interpretation of the sensor data in terms of activity. RBNs detected the wakeup activity accurately with respect to ground truth and their performance exceeded that of an alternative, commercially deployed system, QuietCare.

The Inclusion Enrollment Report does not apply because no activities involving human subjects were performed at any time during the Phase I period of performance.

G. Literature cited

Bishop, C. Living Arrangement Choices of Elderly Singles, Health Care Financing Review, Vol.7, Number 3,

Spring 1986, pp.65-73. Boyen, X. and D. Koller. Exploiting the Architecture of Dynamic Systems, Proceedings of the 16th National

Conference on Artificial Intelligence (AAAI), 1999, pages 313--320. Centers for Disease Control - National Center for Injury Prevention and Control. Falls and Hip Fractures

Among Older Adults, http://www.cdc.gov/ncipc/factsheets/falls.htm, accessed 21 March 2005. Crisan D. and A. Doucet. A Survey of Convergence Results on Particle Filtering Methods for Practitioners.

IEEE transactions on Signal Processing, VoI 50 # 3 March 2002. Gumming, R.G. Epidemiology of medication-related falls and fractures in the elderly. Drugs and Aging

1998;12(1):43-53. D'Ambrosio, B., E. Altendorf, and J. Jorgensen, Probabilistic Relational Models of On-line User Behavior.

WebKDD-2003, Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, WebKdd2003 workshop, Washington, DC, August 24-27, 2003. D'Ambrosio, B. Adaptive Distributed Modeling for Counterspace Fusion and Situation Awareness. In

Proceedings of AIAA 1st Intelligent Systems Technical Conference. Chicago IL, Sept. 2004. Domingos, P and G. Hulten, Mining High-Speed Data Streams. In Proc. Sixth lntl Confernece on Knowledge

Discovery and Data Mining, pp. 71-80, Boston. MA. ACM press, 2000. Doucet, A., N. de Freitas, K. Murphy, S. Russell, Rao-Blackwellised Particle Filtering for Dynamic Bayesian

Networks. Uncertainty in Artificial Intelligence (UAI2000). Dunlop, B., J. Wells and G. Wilensky, The Influence of Source of Insurance Coverage on the Health Care

Utilization Patterns of the Elderly, Journal of Health and Human Resources Administration, Vol. II, Number

3, Winter 1989, Table 3. Friedman, N., L. Getoor, D. Koller, A. Pfeffer. Learning probabilistic relational models. In: Proceedings of the

Sixteenth International Joint Conference on Artificial Intelligence, Morgan Kaufmann Publishers Inc., 1999,

1300-1309 Glascock, A.P. and D. M. Kutzik Behavioral telemedicine: A new approach to the continuous nonintrusive monitoring of activities of daily living. Telemedicine Journal, 6(1):33-44, 2000 Getoor, L., N. Friedman, D. Koller, and B. Taskar, Learning Probabilistic Models of Relational Structure,

Eighteenth International Conference on Machine Learning (ICML), Williams College, June 2001. Graafmans W.C., M. E. Ooms, H.M.A. Hofstee, P.D. Bezemer, L.M. Bouter, P. Lips. Falls in the elderly: a prospective study of risk factors and risk profiles. American Journal of Epidemiology, 1996; 143:1129-36. Gruenberg L. and CP. Tompkins, Including Disability Status in the AAPCC, Waltham, Massachusetts, Health

Policy Center, Brandeis University, 1986. Haddon, W., Jr., E.A. Suchman., and D. Klein, Accident Research Methods and Approaches, Harper & Row,

NY, 1964. Haigh, K.Z. et al. The Independent LifeStyle AssistantTM (I.L.S.A.): Lessons Learned, Honeywell Laboratories

Technical Report ACS-P03-023, 2004. Hanley,R., L. Alecxih, J. Wiener and D. Kenπell, "Predicting Elderly Nursing Home Admissions: Results From the 1982-84 National Long-Term Care Survey," Washington, D.C.: The Brookings Institution, August 1 ,

1989; Teresa A. Coughlin, Timothy D. McBride, and Korbin Liu, "Determinants of Transitory and

Permanent Nursing Home Admissions," Working Paper 3726-03, Washington, D.C., Urban Institute,

January 1989. Hausdorff J.M., D.A. Rios, H. K. Edelber. Gait variability and fall risk in community-living older adults: a 1-year prospective study. Archives of Physical Medicine and Rehabilitation 2001 ;82(8): 1050-6. Hayes, T. L₁ MTPto/&,'PK^*Seft4tlMt,-$rfarKM Λdami. Unobtrusive Monitoring of Health Status in an Aging

Population. In: UbiHealth 2003: The 2nd International Workshop on Ubiquitous Computing for Pervasive

Healthcare Applications Heckerman, D., A tutorial on learning with Bayesian networks. Technical Report MSR-TR-95-06, Microsoft

Research, Redmond, Washington, 1995. Revised June 96 Hooker, K. In: Aging Center to link individuals, families and environments. OSU This Week 44; 21 : 1 , 4.

March 3, 2005. Hornbrook M. C, V.J. Stevens, D.J. Wingfield, J. F. Hollis, M. R. Greenlick, M. G. Ory. Preventing falls among community-dwelling older persons: Results from a randomized trial. Gerontologist 1994:34(1): 16-23. Katz, S., A. B. Ford, R.W. Moskowitz, B.A. Jackson and M.W. Jaffe, Studies of Illness in the Aged, Journal of the American Medical Association, September 21, 1963, pp.94-9. Koller D., Probabilistic relational models. In S. Dzeroski and P. Flach, editors, Proceedings of 9th International

Workshop on Inductive Logic Programming (ILP99) . Springer, 1999. Kutzik, D.M. et al.; US Patent numbers 6,108,685 and 5,692,215. System for generating periodic reports, generating trend analysis and intervention in accordance with trend analysis from a detection subsystem for monitoring daily activity. Lawton, M. P. and E. Brody, Assessment of Older People: Self-Maintaining and Instrumental Activities of Daily

Living, Gerontologist, vol.9 (Autumn 1969), pp.179-86. Liu, K. and K. Manton, "Effects of Medicare's Hospital Prospective Payment System (PPS) on Disabled

Medicare Beneficiaries," Final Report to the Department of Health and Human Services, Office of the

Assistant Secretary for Planning and Evaluation, Washington, D. C, The Urban Institute, February 1988. Lord SR, J. Dayhew. Visual risk factors for falls in older people. Journal of the American Geriatrics Society

2001 ; 49:508-15. Manton, K.G., A Longitudinal Study of Functional Change and Mortality in the United States, Journal of

Gerontology, vol.43, Number 5, September 1988, pp.S153-S161. Marthi, B., H. Pasula, S. Russell, and Y. Peres, Decayed MCMC Filtering. In Proc. UAI-02, Edmonton, Alberta:

Morgan Kaufmann, 2002.

MetLife, The MetLife Market Survey of Assisted Living Costs, October 2003. Milch, B., B. Marthi, and S. Russell, ^"BLOG: Relational Modeling with Unknown Objects." In Proc. ICML-04

Workshop on Statistical Relational Learning, Banff, Canada, 2004. Morris, M. and J. Lundell, Ubiquitous Computing for Cognitive Decline: Findings from Intel's Proactive Health

Research, Intel Research, 2003. Online: http://search.alz.org/Research/Care/lnteLUbiquitousComputing.pdf Neville, J. and D. Jensen. Iterative classification in relational data. In Proc. AAAI-2000 Workshop on Learning

Statistical Models from Relational Data, pages 13--20. AMI Press, 2000. Newman, S., R. Struyk, P. Wright, and M. Rice, Overwhelming Odds: Caregiving and the Risk of

I nstitutio n al ization , (U.I. Report 3691-01 , Washington, D.C., The Urban Institute, November 1988). Pearl, J., Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Morgan Kaufmann,

1988. Perkowitz, M., M. Philipose, K. Fishkin, DJ. Patterson, Mining Models of Human Activities from the Web. In:

Proceedings International WWW Conference, New York, 2004. Philipose, M. et al, Probabilistic Activity Toolkit: Towards Enabling Activity-Aware Computer Interfaces, Intel

Research Seattle Technical Memo IRS-TR-03-013, 2003. Ray W, M. R. Griffin. Prescribed medications and the risk of falling. Topics in Geriatric Rehabilitation, 1990;

5:12-20. Rossignol AM. Enhancing the quality and usefulness of epidemiologic studies of burn injuries. Sums, 22: 429-

432, 1996.

Sanghai, S., P. Domingos, and D. S. Weld, Dynamic Probabilistic Relational models. IJCAI-03, Aug 2003. Savage, LJ. The Foundations of Statistics. Wiley 1954; rev Dover 1972. Takikawa, M., B. D'Ambrosio, and E. Wright, Real-time Inference with Large-Scale Temporal Bayes Nets. In:

Proceedings of the 1$^h Annual Conference on Uncertainty in A! (UAI), Edmonton, Canada, July, 2002. Tapia, M. and E. Intille, Activity recognition in the home setting using simple and ubiquitous sensors. S. S.;

Proc. of Pervasive 2004. Springer (2004). ^"inetti ME, M. Speechley':WeVeMiόn"δf"fal1s among the elderly. New England Journal of Medicine

1989;320(16):1055-9. fan Gelder, S. Health Insurance Association of America, Washington, D.C., in Measuring the Activities of Daily

Living Among the Elderly: A Guide to National Surveys, Joshua M. Wiener and Raymond J. Hanley, The

Brookings Institution, October 1989, prepared for the Forum on Aging-Related Statistics by the Committee on Estimates of Activities of Daily Living in National Surveys.

Claims

CLAIM

1. A method for modeling real-time behavior of a system, the method comprising: providing a dynamical relational Bayesian network; developing an instance network for tracking time with respect to dynamical variables; employing a filtering method to draw inferences from the dynamical relational Bayesian network.

2. The method of claim 1 wherein the system is in-residence patient behavior.

3. The method of claim 1 wherein the system is a collection of moving, physical targets.

4. The method of claim 1 wherein the system is a set of automobiles in traffic.