Download pdf - Functional modeling for enabling adaptive design of devices for new environments

Functional modeling for enabling adaptivedesign of devices for new environments

Sattiraju Prabhakar a,* & Ashok K. Goel b

aSchool of Computing Sciences, University of Technology, PO Box 123, Broadway, Sydney, NSW 2007, AustraliabCollege of Computing, Georgia Institute of Technology, Atlanta, GA 30332, USA

Current AI research on adaptive design is limited to devices with minimal interactionswith their environments. This is partly because current functional models are limitedto devices with minimal environmental interactions. We describe a functionalrepresentation scheme, called the Environmentally-bound Structure–Behavior–Function (or ESBF) model, to represent and organize knowledge of the functioningof a device, including the role of its environmental interactions. We also describe aprocessing strategy called Environmentally-driven Adaptive Modeling (or EAM) formodifying a known design for operation in a new environment, and thus for carryingout new functions that arise due to new device–environment interactions. Weillustrate the ESBF and EAM through the example of adapting the conceptual designof a refrigerator to obtain a preliminary design of an air-conditioner. We also discussthe implementation of this example in a computer program called EnvironmentalKRITIK. q 1998 Elsevier Science Limited. All rights reserved.

Key words:functional modeling, design, adaptation, reuse of experiences, environ-mental interactions.

1 DEVICES AND THEIR ENVIRONMENTS

Much AI research on device design has focussed on deviceswhose functioning can be modeled with minimalrepresentation of environmental interactions. Consider, forexample, classical AI design systems such as R1,1

AIR-CYL,2 PRIDE,3 VEXED4 and VT.5 Typically, thesesystems use heuristic search for a new device design,accommodating interactions among the components of thedevice but largely ignoring the interactions of the devicewith its environment. In this paper, we present an approachfor designing devices that makes use of functional models ofdevices which incorporate, into the model, the interactionsbetween the device and its environment.

1.1 Functional modeling for design

A design task can be defined as transforming the require-ments of an artefact to a physical product description of thatartefact that satisfies the requirements. The designedproducts can also be represented using qualitative modelswith the structure and behavior of the artefact. Physicaldescriptions and qualitative behavioral descriptions cannot

address the tasks required to be performed during theprocess of designing the artefact.

(1) Verification and prediction: once the design task iscomplete, it should be possible to verify the artefactdescription to see whether it satisfies the require-ments or not. It is also required to test, at variousstages of the design process, that the artefactdescription addresses the requirements.

(2) Incorporating designer’s rationale: the artefactdescriptions need to provide explanations as tohow the devices work. This will enable one to testwhether the artefact will work or not.

(3) Addressing the synthetic task: design requiressynthesizing component-related descriptions into anartefact description. This is in addition to analyticaltasks such as simulation of the artefact behavior.

The functional modeling approach suggests that anartefact is modeled in terms of function, i.e. what it isintended to do, and how the intentions are to beaccomplished through causal interactions among com-ponents of the device.6 Here, the requirements are specifiedas functions of the artefact. Gero’s7 method for designingartefacts, using functional modeling, suggests that thestructure of the artefact is not directly generated fromthe requirements. In this model of design synthesis, the

Artificial Intelligence in Engineering12 (1998) 417–444q 1998 Elsevier Science Limited

All rights reserved. Printed in Great Britain0954-1810/98/$19.00PII: S 0 9 5 4 - 1 8 1 0 ( 9 8 ) 0 0 0 0 3 - X

417

*Author to whom correspondence should be addressed.E-mail: [email protected]

expected behavior of causal interactions is generated fromthe function. The expected behavior is transformed to a struc-ture. From the structure, the physical behavior of the artefactis derived. The expected behaviors are compared to the actualbehaviors of the structure of the physical artefact. Theexpected behavior is reformulated based on the comparison.7

The functional modeling work in design has attempted toaddress a variety of design tasks using different kinds offunctional models. These models can be summarized ashaving the following features:8,9

(1) Representational grounding: the functions canbroadly differ based on whether they are theintentions on structure or behavior. The behaviorsof an artefact are abstracted to satisfy a designintent. Here, the functions are grounded in theabstracted behaviors.6,10 An example of groundingfunctions over structure is a display function withthe intent of displaying some aspects of structureof the artefact.11

(2) Levels of intention: the functions vary in the level ofabstraction of function in terms of its meaning.

(3) Functional decomposition: a function of an artefactis decomposed into subfunctions, where each sub-function is mapped onto different aspects of theartefacts.

(4) Relating function with other kinds of knowledge:qualitative behavioral models have used generalfirst principles knowledge to generate behavioralmodels. Functional models represent the specificroles of the first principles in generating behaviorsegments in order to achieve a functionality.6,10

(5) Representing function in context: functionalrepresentations that explicitly embed an artefactinto its environment.12

1.2 Functional modeling for adaptive design

An alternative method to routine design used in designresearch is adaptive design, in which old design cases areadapted to address new design problems.13,14 For adaptingpast design cases to design new devices, qualitative modelsof devices are used to represent design cases.15–18 Thesequalitative models represent relationships between structureand behavior of devices, with minimal representation ofenvironments.

In addition, functional modeling of devices has allowedformalization of the design task and the process of adaptivedesign.13 A design task can be defined as mapping from afunction of the device or the goals of the designer for thedevice to the structure of the device. Each design case isrepresented as a mapping from the structure of the device tothe function of the device. The design of a new device, i.e.achieving a new functionality, is done by adapting astructure to function mapping of a known device to achievethe new functionality. The functional models are used torepresent design cases, from which adaptation spaces aregenerated for solving a new design problem. The adaptation

spaces for designing new devices can be generated provideda functional model has the following properties.

• Localization of functional aspects of the deviceonto structural aspects of a design case. Thislocalizability property is not inherent in allmodels, e.g., models of non-linear functions.

• Isolable modificationof a localized element of apast design case. The modification of local elementsshould preserve the remaining models of the designcase, primarily the localizability property of the restof the model. This enables the model to preserve theremaining functional aspects and their adaptability.A further consequence of this property is that theadapted model itself is adaptable, thus ensuring apotentially infinite sequence of adaptable models.Therefore, the adaptation spaces generated containmodels that are functional variations of the pastdesign.

The Structure, Behavior and Function (SBF) model10

has the above properties. KRITIK10,16 and IDEAL19,20

demonstrate the usefulness of these properties in a numberof systems for various domains.

The following aspects of the SBF models enablelocalization and isolable modificationproperties to besatisfied.

(1) Functional closure of behavior: in SBF models, onlyfunctionally relevant behaviors are represented. Thisallows functions to be directly mapped onto behaviors,and no other behaviors are present in the functionalmodel that does not have functional significance.

(2) Causal closure over structure: each of the behavioralstates in an SBF model is closed over a device sub-structure and includes a small set of structuralelements of the device. A causal event links twosuch states by a behavioral state transition. Thisaspect of an SBF model suggests that if there is afunctionally relevant event, then it can be mappedonto a small group of structural elements.

(3) Equivalence of events across functions: two eventsare equivalent if their behavioral states andbehavioral state transitions are equivalent. Thestates, or the state transitions, are equivalent iftheir elements are equivalent. Both of these arerepresented using a global ontology that ensurestheir equivalence. This point enables modifiabilitywithout disturbing the equivalence structure in therest of the model.

(4) Localizable causes: a state is a causal result ofanother state that lies in a sequence of state transi-tions within the same behavior. This makes use ofthe transitive nature of the causal relationshipsbetween states. This closedness makes the searchfor a behavioral state transition localizable.

Many functional models possess one or two of the aboveproperties,21–23but not all. SBF models provide a powerfulsolution for adaptation problems.

418 S. Prabhakar, A. K. Goel

However, this functional modeling of devices allows onlya minimal modeling of the environment:

(1) Minimal role of the environment: the device isassumed to have a small set of input and outputstates compared to its large number of internalbehavioral states. Changes in behavioral states ofenvironment often do not contribute to thefunctionality of the device.

(2) Passive role of the environment: changes in thestates of the environment are brought about by thedevice, but the environment plays no distinct role inthe functioning of the device.

For these reasons, SBF models are limited in designingdevices for new environments. Furthermore, they face thefollowing problems while addressing the design for newenvironments by adapting past experiences.

(1) Incomplete model generation: since the environmentof a device is not explicitly modeled, the adaptationof SBF models generates incomplete models—amodel of the device that does not capture the aspectsof its environment.

(2) Non-adaptable adaptations: multiple SBF modelsdescribing different designs may be composed toarrive at an SBF model of the new device.16 Sinceeach SBF model represents environmental inter-actions, however minimally, the resulting SBFmodel would also minimally represent the environ-mental interaction. Since the environmental inter-actions are not explicitly modeled in an SBFmodel, this combination of SBF models in anadaptation can give rise to unpredictable results.That is, it is not possible to conclude which environ-mental aspects the adapted SBF model wouldincorporate.

(3) Large adaptation spaces: since an adaptation doesnot consider the environments, it is underconstrainedfor the requirements of the environments. The result-ing adaptation space may be very large orinappropriate for an environment.

The following issues become important in modelingenvironments for adaptation.

(1) Easy replaceability of environments: devices arephysically moved from one environment to another,but can still deliver their functions. The functionalmodel of the device needs to meet this criterion.Since a designer may not be aware of all environ-ments in which a device can be operated or theremay be no generic model of all environments, thedevice models will often be incomplete. The onlysolution is to change the model of the environment.This change in environmental model may suggestsome changes in device model, but together theystill need to support a consistent behavior.

(2) Supporting close, yet acausal, interactions betweendevice and environment: an environment is not a part

of a device in the same sense that a component of adevice is a part. This is because the environmentdoes not causally contribute to the functionality ofthe device. An environment often provides aphysical situation in which the causal events of thedevice take place. Devices and their environmentschange dynamically and their event structuresdepend upon each other. The functional modelshould be able to capture this dependency.

(3) Modeling environments without intrinsic functions:in contrast to devices which are designed for specificfunctions, environments are often not designed tosatisfy any functions. For example, the environmentof an automatic coffee-maker can be a room which isnot specifically designed to support the function ofcoffee-making. Devices are said to haveintrinsicfunctions, whereas environments give rise toascribed functions. The inherent process of thedevice supports the intrinsic functions. On theother hand, the ascribed functions are achieved byascribing a structure to the environment that cansupport the function.

In order to address the above issues, we have developed anew functional representation scheme that extends the SBFmodel to represent device–environment interactions andtheir role in the functioning of the device. We call the newscheme the Environmentally-bound Structure–Behavior–Function (ESBF) model. In contrast to SBF models andother functional representation schemes, the ESBF modelviews a device function not as an abstraction of the internalbehaviors of the device, but as an abstraction of the inter-actions of the device with its external environment. ESBFgives rise to a new processing strategy, called Environmen-tally-driven Adaptive Modeling (EAM), for adapting thedesign of a known design for operation in a new environ-ment and for carrying out new functions based on the newenvironmental interactions. Unlike the adaptive modelingstrategy used in the KRITIK system, EAM explicitly takesinto account device–environmental interactions.

ESBF and EAM arise from a detailed analysis of threehistorical case studies of technological invention: theautomatic coffee-maker,24,25 the room air-conditioner26,27

and the windmill.28 In this paper, we illustrate the ESBFmodel and the EAM strategy through the example ofevolution of the preliminary design of the air-conditionerfrom the conceptual design of the refrigerator.29,30 Thediscussion focusses on the adaptation task as opposed toother tasks of adaptive design, such as design retrieval,design verification, redesign and design storage.

A computer program called Environmental KRITIK(E-KRITIK) implements and evaluates the EAMstrategy for the task of adapting an ESBF of therefrigerator into an ESBF model of the air-conditioner.In this paper, we briefly describe E-KRITIK. Wealso discuss its limitations and how it might beextended for addressing some of the other tasks of adaptivedesign.

Adaptive design of devices 419

2 ADAPTIVE DESIGN OF DEVICES FOR NEWENVIRONMENTS

The functions of many engineered devices depend upon theinteraction between the devices (made up of componentsand connections designed by the designer) and theirenvironments. For example, a buzzer circuit makes a buzz-ing sound within air by producing mechanical vibrationswithin the air surrounding it. A buzzer circuit designed foran ‘air only’ environment may not produce sound in anenvironment filled with a magnetic field. This is due tointeractions between the magnetic field of the magneticcoil of the buzzer circuit and the environmental magneticfield. The model of the buzzer circuit, which operates withina magnetic field, can explain how the device operates byrepresenting the interactions between the device and theenvironmental magnetic field.

In this model, the function is a state change in theenvironment that is brought together by the device and theenvironment by continual interaction. The function is a setof designer’s goals where each goal is instantiated with abehavioral state of the environment of the device.

2.1 The ESBF representation

In order to address the above issues, we propose a newmodel (ESBF) which is an extension of the Structure,Behavior and Function (SBF) model.10 The ESBF modelsupports a hierarchy of interactions between the deviceand its environment and is illustrated in Fig. 1. Before weproceed further, we introduce a few terms in order to avoidambiguity. We call the model of a device thedevice model,the model of the environment theenvironment model, andthe model that includes both device and environmentmodels theintegrated device model.

We list below the features of this new models and thenrelate them to the issues we identified earlier.

(1) The environment is considered as a ‘device’ whichsupports the device model, so that the integrateddevice model achieves its function. The environmentmodel has anascribed behaviorto support devicemodel functionality. For example, the flying functionof an airplane can be achieved by the plane having

the function of thrust generation, whereas the air hasthe behavior of counter-thrust generation. Only thoseenvironmental processes are modeled that enable themodel to support the functionality of the device. Theinherent structure, as opposed to the ascribedstructure, of a device and its interactions with theenvironment support itsintrinsic functionality. Thedevice model captures the device’s ability to makeactive changes in the environment and react tothe environmental changes as contributing to theintrinsic functionality. The explicit modeling of theenvironment addresses two issues of designing fornew environments: modeling and prediction. Sinceenvironment models are modular, they can bereplaced by other environment models. Thisaddresses the easy replaceability issue of modelingthe environment.

(2) The interactions between various structural elementsof the device and the processes of the environment’sphysical system are organized into a hierarchy.While achieving the functionality of the device, theprocesses of the environment model interactmore closely compared with interactions betweenelements belonging to structural elements of deviceand environment models. Due to this organization ofinteractions, it is possible to localize the functionalaspects onto structural aspects in a principledmanner.

(3) Each of the device and environment models followthe modeling principles of an extension of SBFmodels. Thus, adaptation can be addressed withineach of these models by functional closure ofbehaviors, causal closure of structure and localizablecauses. The event structures of environment anddevice models are not causally related. They arerelated by influencing the state transitions of causalbehaviors. This allows the adaptation strategies tolocalize the structural elements for functional aspectswhile capturing various levels of interactions thattake place with the complex system of a deviceand its environment.

(4) The behavioral states and state transitions inboth environment and device models sharemany ontologies: substance, component (location),

Fig. 1. The features of the ESBF model.


property and parameter. This allows isolablemodifiability in adaptation.

The following is a list of features of the SBF model.10,16

(1) Functions are represented as transitions of behavioralstates.

(2) Internal causal behaviors of devices are explicitlyrepresented to explain how the function of a deviceis delivered by the structure of the device.

(3) Compositions of causal behaviors are alsorepresented in delivering the function of the device.

(4) The causal and compositional aspects of behaviorsrepresent the interactions among the structuralelements. The structural elements are broadlycategorized into components, substances and theconnections between the components.

(5) Changes to structural elements are represented bybehavioral state transitions.

(6) A set of behavioral principles explains thebehavioral state transitions. The basic set ofbehavioral principles is: (i) applying a physicalprinciple; (ii) using a function of a component; (iii)using a structural relationship; (iv) causaldependence on another behavior.

The ESBF model has the following specific aspects (seeFig. 1).

(1) The function of the integrated device, the functionalabstraction of the device and the behavioralabstraction of the environment are represented asbehavioral state transitions.

(2) The behaviors of the device and environment canbe of two types: causal and compositional. Com-positional behaviors extend the causal behavior ofan SBF model by allowing more than one causalsegment to give rise to one single behavioral state,and a single behavioral state causing multiple causalbehaviors.

(3) The models of a device and its environment mayhave classes of structural elements that are quitedifferent. For example, a device may have largelumped structural elements, whereas the environ-ment may be modeled as having elements whichare at a lower level of grain size. These differencesin the grain sizes of the structural elements mayrequire different behavioral principles.

(4) A new ontological entity is added to the environmentbehavioral states: the qualitative region. Aqualitative region need not be a physical componentwith an intrinsic functionality, but a region ofphysical space that shows a behavior. For example,the surface of a wall deflects an air-current hitting itat an angle.

(5) The causal state transitions of behaviors ofboth the device and the environment can becausally dependent upon each other. This isrepresented by a new behavioral principle calledEnvironment_Device_Interaction.

2.2 The EAM process

Before we present an adaptation strategy that addressesthese issues, let us define the adaptation task for the designof new environments.

Fig. 2 illustrates the overall adaptation strategy. Given afunctional specification, Fn2, of a new integrated deviceID2, a known model M1 of an integrated device ID1 withfunction Fn1 is retrieved from the case base of designs bymatching the functionalities Fn1 and Fn2. The functionaldecomposition that is present in model M1 for its deviceand environment is adapted to arrive at the functional andbehavioral abstractions of device and environment of ID2.In model M1, the device model is MD1 and the environmentmodel is ME1. In MD1, functional abstraction differencesbetween the device D1 of ID1 and the device D2 of ID2 arelocalized to structural elements. In ME1, the behavioralabstraction differences between the environment E1 ofID1 and the environment E2 of ID2 are localized to someprocesses.

These models specify the adaptation spaces. Whilesearching these adaptation spaces, constraints are generatedthat reflect the modeling choices which are passed on to themodification processes in the rest of the models that arebeing adapted. Each of the modified device and environ-ment models satisfies the generated constraints.

From each of the adaptation spaces a model is selected,and the models are composed to form a complete model ofthe integrated device ID2. The localization and modificationprocesses within the device and the environment modelsare different. In this paper, we focus primarily upon theadaptation strategy of the environment.

3 THE ESBF REPRESENTATION: A DETAILEDEXAMPLE

We present an ESBF model for the integrated devicerefrigerator. Fig. 3 shows a schematic for a refrigerator.30

The purpose of the refrigerator is to pump heat from achamber, where the food is stored, into the surroundingroom. The device of the refrigerator is the part containingthe components: compressor, condenser and evaporator.This device of the refrigerator can be viewed as havingtwo environments: the chamber along with its contentsand the room outside the refrigerator.

The chamber or the enclosure, along with its contents, isan environment for the refrigerator device, because many ofits aspects such as the contents are not designed for thedevice, but can influence the functionality of the device.In other words, the device model can be described mainlyusing endogenous variables and the environment model


using exogenous variables. In this paper, we focus mainlyupon one environment: the chamber. The outside roomenvironment can be considered as a heat sink, without losingthe generality, which provides room temperature as areference point.

The function of the refrigerator—to remove heat from thechamber to the surrounding room—is achieved by a‘working fluid’ or a ‘refrigerant’ going through a cycle ofliquefaction and gasification. Four properties of the ‘work-ing fluid’ change to achieve this function: heat content,pressure, temperature and physical state. The workingfluid enters the compressor as low pressure gas at a hightemperature. The compressor applies pressure on this gas,thus increasing its pressure and temperature. This gas thenflows through the condenser, gives out the heat to the out-side room, cools down to a low temperature and thencondenses into liquid. The liquid will still be at high

Fig. 2. Overall strategy for adaptation of a functional model of an integrated device.

Fig. 3. A schematic of a refrigerator.


pressure. It then enters the evaporator, which is a narrowtube. During the flow through the evaporator, the liquidabsorbs the heat from the chamber and becomes gas atlow pressure. This low-pressure gas at a higher temperatureenters the compressor and the whole cycle repeats.

The purpose of the enclosure is to transfer the heat fromthe food stuff held in the chamber to the evaporator. Thisheat transfer is performed by the air-currents within thechamber. A temperature gradient exists within the enclosuredue to temperature differences between the food stuff andthe evaporator surface. The air-current, after receiving heatfrom the food stuff, moves towards the lower temperaturespatial point immediately next to it. In this case, the lowertemperature spatial point is above the current location.Therefore, the air-current moves in the upwards direction.When it reaches the surface of the evaporator, it exchangesits heat with the surface and the cooler air-current descends.This cycle repeats, thus cooling down the food stuff.

3.1 Functional abstractions of the refrigerator

The function of the refrigerator is given below. Here Hs. Lsand Hh. Lh. Let Lr be the room temperature, then Ls¼ ,Lr. From here onwards, we refer to the food stuff assource.

The behavioral abstraction of the enclosure, given below,plays a supportive role in letting the refrigerator deliver itsfunction. The enclosure or the environment transfers heatfrom the source to the evaporator.

The functional abstraction for the refrigerator device is givenbelow. Here, the heat content Ql1 does not include the heat(Hh ¹ Lh) gained from the environment.

The behavior of the ESBF model of a refrigerator has twodistinct behaviors belonging to its environment and device.The device behavior represents how the heat gained fromthe environment is transferred to the outside room. Theenvironment behavior represents how the heat of the sourceis transferred to the device. The environment behavior is acomposition of two causal behaviors. The first is tem-perature and heat variations of the source due to its inter-actions with the air-currents. The second is the movement,and heat and temperature variations, of air-currents withinthe enclosure, which carry heat from the source to theevaporator of the device.

3.1.1 Function of the refrigeratorFig. 4 illustrates the function of the refrigerator. The givenstate describes the source as having the temperature Hs and themakes state describes the source as having the temperature Ls,with the condenser having the heat quantity Ql1. This functionshows that the source temperature reduces from Hs to Ls,while the heat quantity is moved to the condenser.

3.2 Behaviors of the refrigerator

Figs 5–7 illustrate the detailed SBF model of therefrigerator device. The function of the refrigerator deviceis represented using the location (loc), substance (sub),property (prop) and parameter (param) ontologies. Thegiven state describes that the location, EVAPORATOR,has the substance, REFRIGERANT, which contains thesubstance, HEAT. The REFRIGERANT has the propertiesphysical state (shown as state), temperature and pressure,whereas the HEAT has the property quantity. The functiondescribes the variation of values in these propertiesfrom given to makes states. For the refrigerator device, thefunctional abstraction represents loss of heat (He¹ Le) andreduction in temperature of the refrigerant.

Fig. 6 illustrates the refrigerator device behavior thatachieves its functional abstraction. The two segments ofbehavior illustrated in Fig. 6 belong to the same behavior:the right segment follows immediately after the leftsegment. The first state of the behavior is at the top left,which shows that the refrigerant is at the evaporator in aliquid state with pressure Ph, temperature Le and heat Ql1.The first state transition represents an interaction of theevaporator with the environmental behavior of heattransfer from the source. The evaporator interacts with theenvironment and absorbs the heat (see Fig. 7) from it. Thenext state represents the result of such an interaction: thetemperature and the heat content of the refrigerant are nowhigher, the pressure lower and it is in a gaseous state. Thisinteraction is facilitated by the domain principles: theZeroth Law of Thermodynamics and the Law of GaseousAbsorption of Heat, and a structural relation that theevaporator and air-current are in contact.

The second state transition represents the movement ofthe refrigerant from the evaporator to the compressor due tothe ALLOW function of the pipe that connects them. Thethird state transition represents the compression of thegaseous substances, resulting in a state where the refrigerantis still in the gaseous state with higher pressure, temperatureand heat content. The fourth state transition represents themovement of the refrigerant to the condenser due to thefunction ALLOW of the pipe connecting the compressorand the condenser. Here, a set of behavioral transitionannotations describe the heat loss to the outside room.The cooler and liquid refrigerant moves to the evaporatorusing the ALLOW function of the pipe connecting theevaporator and the condenser.

Fig. 7 illustrates the state transitions that representchanges in the evaporator as it interacts with the environ-ment. Here, only the changes in evaporator temperature are


shown. The changes in the rest of the properties—pressure,heat content and physical state—as illustrated in the states ofFig. 6, can be represented in a similar way. The evaporatorgains heat from the refrigerator enclosure by interactingwith the air-currents. Each air-current can have a differenteffect upon the heat exchange. That is, the amount of tem-perature changed in an evaporator state depends upon thetype of air-current. As will be described later, this iscrucial in representing how the refrigerator works. InFig. 6, each segment shows an interaction. The arrowindicating environment–device interaction illustrates thedirection of heat flow. The loop on the arrows is the locationof the interaction. All these interactions are aggregatedinto a single evaporator state. The first and last statesof this behavior form the first two states in the behavior ofFig. 6.

3.3 Behaviors of the refrigerator environment

The behavioral abstraction of the refrigerator enclosure isfacilitated by the air-currents, which carry the heat from thesource to the surface of the evaporator. This transfer of heatby the air-currents, away from the source to the evaporator,can be represented in four scenarios, which are integratedinto a single representation in Fig. 8. Fig. 8 illustrates fourair-currents: C1, C2, C3 and C4. Each of these air-currentsshows different behavior in carrying heat from the source tothe evaporator and hence realizing the behavioral abstrac-tion of the environment in different ways. To understandthese behaviors further, we need to define a few terms.

In order to capture the behavior of multiple fluid currentsthat are often present in an environment, we introduce a newontology: the qualitative region. In a refrigerator environ-ment, the fluid current is the air-current. An air-current isconsidered here as a substance that brings about changes tovarious aspects of the environment and the interfacebetween the device and the environment. The behavior ofan environment, in its SBF model, captures specific

interactions of air-currents. These interactions occur atqualitative regionsL1, L12, L2, L14 and L13.

Each of these qualitative regions is characterized by itsqualitative location and a component elementat thatlocation. Two spatial locations inside an environment areconsidered as belonging to the same qualitative location ifthey share similar spatial properties, such as on_the_wall,on_the_ceiling, in_the_free_space, etc. The spatial regionsinside the environment, which have some unique qualitativeproperties, are calledcomponent elements. Component ele-ments allow for the localization onto a small spatial regionfor comparison with other regions, in an otherwise con-tinuous space. For example, two areas on a wall can bethe same type of component element if they can have thesame qualitative values for temperature. The qualitativeregions identified in Fig. 8 are further described below.

• Location L1: here, the component element is thesource and the qualitative region is near the floor.

• Location L12: this is the qualitative region on thewall close to the middle section of the environment,and has the component element that is a region onthe wall.

• Location L13: this is a qualitative region on the wallcloser to the ceiling. It has the component elementthat is a region of the wall.

• Location L14: this is a region of free space insidethe environment. The component element is aspatial region.

We can now describe the air-currents in detail. Thesimplest of all the air-current behaviors is that of C1,which is called Vertical Heat Transfer Behavior (VHTB).The air-current C1 exchanges heat with the componentelement source at L1. Since it is hotter, it ascends towardsthe cooler ceiling. It hits the ceiling and exchanges heatthere and, being cooler, descends towards L1. This cyclerepeats itself.

Since there can be more than one air-current present at a

Fig. 4. The function of the refrigerator.

Fig. 5. The functional abstraction of the refrigerator device.


Fig. 6. The behavior of the refrigerator device.

Fig. 7. Device state transitions as a result of interactions with the environment.


time and they can have different temperatures, they caninfluence each other’s behavior. Air-current C2, being oflower temperature compared to C1, is pushed aside by C1.It moves upwards at an angle and, as a result, hits the wall atL12. Here, the wall being at the same temperature as the air-current, no heat exchange occurs. However, the componentelement wall at L12 has the property to deflect the air-current. The air-current now moves upwards at a newangle towards the cooler ceiling. At the ceiling, it exchangesheat with the evaporator and, being cool, descends verticallydownwards. This behavior of air-current C2 is called Walland Air-Current Collision Behavior (WACB). Current C3shows similar behavior.

The air-current C4 starts vertically upwards afterexchanging heat with the source. On its way, at L14, itcollides with a cooler air-current descending vertically.The result can be a vertically downward or upward movingair-current, depending upon the temperature of the resultingair-current. This air-current behavior is called Air-Currentand Air-Current Collision Behavior (ACACB).

3.3.1 Functional abstraction of the refrigeratorenvironmentFig. 9 illustrates the behavioral abstraction of the refrigera-tor environment. The transfer of heat to the device occursimplicitly, by the air-currents, as has been explained earlier.Each state of the function is described using the followingontologies: qualitative location (loc), component element(comp), substance (sub), property (prop) and parameter(param). The function for the refrigerator environmentdescribes the decrement in temperature of the source.

3.3.2 Heat transfer from the sourceFig. 10 illustrates the Heat Transfer Behavior (HTB) of thesource in the environment of the refrigerator. This behaviorrepresents the cooling of the source by three different air-

currents. The three air-current behaviors—VHTB, WACBand ACACB—interact with three behavior segments ofsource cooling. Each of these interactions occurs at abehavioral state transition of the source cooling. Or, thesource cooling is enabled by the exchange of heat, whichis represented as a conditional transition. Here, the arrows ofconditional transition represent the direction of the heattransfer. All the behavioral segments for source coolingare aggregated into a single behavioral state, whichrepresents the aggregated value of the temperature cooling.

The behaviors of the environment have behavior statetransitions which use the annotations of the behaviormodel in KRITIK plus two other annotations. The first isthe environment–device interaction. This is also used indevice behaviors. The second is the qualitative parametricrelation. This represents a qualitative relation between thevalues of the local qualitative parameters. If this relationis true, in addition to the fact that other annotations of thestate transition are also true, then the state transition takesplace.

3.3.3 Vertical heat transfer by air-currentsFig. 11 illustrates the behavior of air-current C1 of Fig. 8.The behavioral state uses the same ontologies as thebehavioral abstraction of the environment. The first stateillustrates that the substance AIR-CURRENT and thecomponent EVAPORATOR are both at the location LCe.The AIR-CURRENT has the properties TEMPERATUREwith value Htac1 and DIRECTION of movement UP. TheAIR-CURRENT has a substance HEAT, which has theproperty QUANTITY with a value of Hhac1.

These property values change, in the first state transition,due to the interaction of the air-current with the evaporatorheating behavior (see Fig. 7), as represented by an environ-ment–device interaction. Due to the first state transition, theair-current loses heat to the evaporator, resulting in the nextstate with lower temperature and heat content. The secondstate transition represents the movement of the air-current inthe DOWN direction. This state transition is enabled by adomain principle that the cooler air-currents descend, anddue to the parametric relation that the distance between theevaporator and the source is less than the temperaturedifference between the source and the air-current. Theresult is that the air-current reaches the location of thesource.

The third state transition represents the heat exchangebetween the cooler air-current and the hotter source. Thisinteraction between the behaviors of the air-current and thebehavior of the source (see Fig. 10) is due to conditionaltransition.

The resulting hotter air-current rises to the evaporator.

Fig. 8. Different kinds of air-currents transporting heat from thesource to the evaporator.

Fig. 9. The function of the refrigerator environment.


This is represented in the last state transition, which has adomain principle for the ascent of hot air-currents and aparametric relation that expresses a condition for ascent ofthe hot air-currents.

3.3.4 Collisions between air-currents and wallsFig. 12 illustrates the behavior of the air-current C2 of Fig. 8.The first state (state 1) shows that the component SOURCEand the AIR-CURRENT (AC) are both at location LCs.The SOURCE has the property TEMPERATURE withvalue Hs2. The AIR-CURRENT has the propertiesTEMPERATURE with value Ltac3 and DIRECTION withvalue DOWN. The AIR-CURRENT has the substanceHEAT, which has the property QUANTITY with a valueLhac3.

This state is changed to a state where the air-current(AC2) has higher temperature and heat content, andthe SOURCE has less. This occurs due to the conditionaltransition with the behavior of source (HTB; see Fig. 10).The hotter air-current (AC2) raises LEFT-UP due to thepresence of a hotter air-current (AC1) close by. This isrepresented in the second state transition, which uses adomain principle that hotter air-currents ascend. It alsouses a parametric relation that the difference in distancesbetween the wall and the source is less than the temperaturedifference between the air-current and the wall. The air-current (AC2) then hits the left wall on its way and getsdeflected towards the ceiling. This behavior is expressedby the third state transition. The last state transitionrepresents the environment–device interaction betweenthe air-current cooling and the evaporator heating. Thecooler air-current (AC2) descends.

3.3.5 Collisions among air-currentsFigs 13 and 14 illustrate the behavior of the air-current C4 ofFig. 8. Fig. 13 illustrates two behavior segments for two air-

currents (AC1 and AC2) that collide at location LC14. Onebehavior segment represents the behavior of an air-currentAC1 that descends after interacting with evaporator heatingbehavior (see Fig. 7). The second behavior segmentrepresents the behavior of an air-current (AC2) thatbecomes hotter by interacting with the source coolingbehavior (see Fig. 10). These two causal behaviors arecomposed into a single behavioral state, the compositionalstate.

Fig. 14 illustrates the behavior that represents the result ofcollision between two air-currents. Here, two behaviors canemerge from collision, depending upon the resultingtemperature of the air-currents. The resulting properties ofair-currents can be computed using a number of parametricrelationships between the local values of the qualitativeproperties. From these relationships, it is possible todecide the direction of an air-current. If a cooler air-current,compared to source, results from the collision, then the air-current moves downwards. On the other hand, if a hotter air-current, compared to the evaporator, results from collisionthen the resulting air-current moves upwards. As they reachthese locations, heat exchange again takes place by using theconditional transition and environment–device interactionbehaviors.

4 THE EAM PROCESS: DETAILED ALGORITHMS

In Section 2, we considered an overall strategy for theadaptation in Environmental Adaptive Modeling (EAM).This section focusses primarily upon the strategy for theadaptation of environmental models, with a brief descriptionof device model adaptation. These two independentstrategies can produce an integrated device model due toexchange of constraints during the adaptation process.

Fig. 10. Environment source behavior for interaction with different air-currents.


This section demonstrates the environmental modeladaptation strategy by adapting a refrigerator ESBF model,to arrive at an air-conditioner ESBF model. The problem indesigning an air-conditioner is specified in terms of itsfunction and also the structural description of theenvironment in which it works. These are given below.

Here, Hhn and Lhn are the sets of initial and finalheat contents of then heat sources. Hhn–Lhn representsa set of heat quantities pumped into the room outsidethe air-conditioner. Tp and Tf are the set of temperaturesof n individual sources. Tf, Tp indicates that everysource is at a lower temperature in themakes stateof the function compared to itsgiven state. We choosea problem for which the environment structure hasthe following properties: hke, hne, lke , lne

and wke, wne, where hke, lke and wke are height,length and width of the refrigerator environment,respectively.

The adaptation strategy, as illustrated in Fig. 2, takes thefunctional abstraction of the refrigerator device andbehavioral abstraction of its environment, then adaptsthem to the air-conditioner device and environment,respectively. Thus, this strategy posits a functionaldecomposition for the air-conditioner similar to that of therefrigerator.

The new behavioral abstraction for the air-conditionerenvironment is adapted from the functional abstraction ofthe refrigerator environment (see Fig. 9). The generatedbehavioral abstraction for the air-conditioner environmentis given below. Here, the temperatures ofn heat sources arereduced from Hs to Ls. This strategy for function generationis discussed in detail later.

The functional abstraction of the air-conditioner device isgiven below. Here, the air-conditioner device takes the heatcontent Hhn¹ Lhn that is transported fromn heat sources atits input, to the outside room. The resulting heat content ofthe condenser refrigerant is Ql2.

Having come up with these functional abstractions, thebehavior and the structure of the device and environmentfor the air-conditioner are designed to satisfy thesefunctional abstractions.

4.1 Adapting the ESBF model

In order to arrive at the functional model of the air-conditioner, the functional model of its environment,which supports the function of the integrated air-conditionerdevice and environment, is required. The adaptationstrategy to arrive at the functional model of the air-conditioner environment, from the refrigerator environmentmodel, is different from the adaptation of the device model.In device model adaptation, the differences between theknown device function (e.g., refrigerator device) and thenew device function (e.g., air-conditioner device) are loca-lized onto the behavior of the known device. Then, theknown behavior model (e.g., the refrigerator behavior

Fig. 11. The vertical heat transfer behavior of air-current C1.


Fig. 12. The wall and air-current collision behavior of air-current C2.

Fig. 13. The collision of two air-current behaviors of C4.


model; see Figs 6 and 7) is modified such that it can deliverthe new device function.10,16

In environment adaptation, the differences between thefunctions of the known and new environments are minimal.The environment adaptation strategy is different from thedevice adaptation strategy for the following reasons.

(1) The structure of the new environment is alreadygiven in the design problem. The design process isrequired to arrive at the behavioral model of theenvironment without modifying the structure of thenew environment.

(2) The environmental behavioral model requires asubstantial amount of spatial reasoning. Often,fluids (e.g., air) or fields (e.g., magnetic field)cause interactions between the structural elementsof the environment to occur. Identifying changesand their effects, due to the fluids or fields, requiresspatial reasoning.

(3) Since the device SBF models are built upon lumpedelement models of the devices, replacing a set oflumped elements with another set of lumped ele-ments is a useful adaptation strategy.10,16,20In envir-onmental adaptation, formulating such elements,where useful interactions can take place, is a majoraspect of adaptation strategy.

(4) The design process can change only one parameterfor modifying the physical aspects of the environ-ment: inputs to the environment. This modification isrequired for the environmental model to achieve itsbehavioral abstraction.

The following is a description of the task for environmentalmodel adaptation.

4.1.1 The strategyIt is to be noted that the primary purpose of adaptation of aknown environment model to a new environment function isnot to develop a complete simulatory model of the environ-ment. Rather, it is to address the question ‘In order to satisfythe behavioral abstraction of the new environment, whichinputs and outputs should its model have?’. That is, for theair-conditioner environment, in order for it to transfer theheat from its sources inside the enclosed room to theevaporator surface, the adaptation strategy needs to identifyits inputs and outputs. The requirements for the environmentto have these inputs and outputs are formulated asconstraints. This constraint discovery process and itssatisfaction is discussed in detail later.

The question addressed by the environment adaptation isrelated to the objective of adapting a functional model of thedevice: ‘To arrive at a model of a device that responds to theoutputs from the model of its environment such that it candeliver the function of the integrated device and environ-ment’. For example, the air-conditioner device should takethe heat supplied by its environment and then transfer it tothe outside room.

This formulation of the environment model conformswith the original conception of behavioral abstraction,presented in Section 2, for the environment as ascribedand that it supports the device to achieve the functionalityidentified for the integrated device and environment.

Our strategy, in order to address this question, is tosimulate the environment to the extent that it can bevalidated against its identified behavioral abstraction. Ifthis limited simulation model of the environment fails todeliver its behavioral abstraction, the cause for failure isdiagnosed. Since the structure of the environment cannotbe modified, this cause is posed as a constraint on the devicemodel. The device model is revised in order to eliminate thecause of the failure in the environment model, by generatinga new input to the environment, while still maintaining itsfunctionality. The environment model is enhanced with thechange and the simulation is resumed to identify furtherproblems with satisfying the behavioral abstraction of theenvironment. This process repeats until there are no moreproblems with the environment model in satisfying itsbehavioral abstraction.

One of the major problems in simulating a largeenvironment is its complexity. For example, in a refrigeratorenvironment, it has been possible to explain the behavior ofits environment by identifying four qualitatively differentair-currents. In a larger environment with several heatsources, the number of air-currents can be hundreds. Insuch cases, simulation is not computationally viable.Furthermore, in such complex models, it may not bepossible to identify new criteria that a behavioral model ofthe environment completely satisfies in its behavioralabstraction.

In order to address these two issues, the method suggestedhere is a heuristic-driven adaptation strategy. In the lattersections, several general strategies are presented that useheuristics as part of this adaptation strategy. For the domainwhere the environmental interactions are caused by the fluidcurrents, and the fluid currents by a source, the followingtwo heuristics address the above two issues for environmentadaptation. These heuristics are available to variousstrategies that adapt the ESBF models and are applied auto-matically by the strategies.

(1) Analysis of Structural Patterns (H1): we view thenew environment behavior as reproducing theknown environment behavior in some sense. Ifthe new environment has structural features thatare multiples of the known environment, then thebehavior is also assumed to multiply. If the structuralfeatures are a fraction of the known environment


features, then the known environment behaviors areconstrained. Due to this view, it is possible to selecta representative set of repeatable behaviors in thenew environment model and predict from it the fea-tures relevant to the behavioral abstraction of thenew environment. This eliminates the need todevelop the complete model of the new environment.An example of a representative set of repeatablebehaviors of the refrigerator environment model forthe air-conditioner is a set of three refrigeratorbehaviors (see Fig. 7) adjacent to each other. Wewill study this model in detail later.

(2) Analysis of Behavioral Patterns (H2): a comparativeanalysis of four air-currents in refrigerator environ-ment behavior (Fig. 8) shows that the vertical heattransfer behavior of air-current C1 is the mosteffective in satisfying the behavioral abstraction for

the refrigerator environment. This is because C1does not incorporate collisions with other air-currents and walls. Collisions are the means bywhich the heat is lost or the transportation of heatis delayed. The functionality of an environment canbe drastically affected if the vertical air-currents arelost in an environment.

Using theAnalysis of Structural Patternsheuristic, it ispossible to eliminate the complexity problem of simulatinglarge environments. TheAnalysis of Behavioral Patternsheuristic can identify when an environment model has failedto achieve its behavioral abstraction. Fig. 15 illustrates theoverall strategy for the adaptation of the environment modelfrom a known environment model.

This subsection briefly presents the overall strategy forthe adaptation of the environment model, and each of the

Fig. 14. The behaviors resulting from a collision of C4.


individual steps is further explained in the subsequentsubsections. Thefunction generationproduces a behavioralabstraction of the new environment from the behavioralabstraction of the known environment, and the structuresof the known and new environments.

In thecompositionstep, the new and known structures ofenvironments are compared for their structural differences.This is to see whether the new structure can be divided intoregions where each region resembles the known environ-ment. A region of the new environment resembles thestructure of the known environment if it has a source andhas the same dimensions as the known environment. Theregions of the new environment structure that resemble theknown environment structure are grouped to a set R1. Allother regions of the new environment structure are groupedto a set R2. This process of dividing Sne into regions of Skeis calledregionalization.

Into a selected set of R1 regions, Rs, identified by theAnalysis of Structural Patterns(H1) heuristic, the behaviormodel of the known environment is transferred. The resultof this transferinto Rs results in a partial behavior model forthe new environment, Bne.

Extensionsare made on the behavior of Bne over theregions of R2 such that aspects of Sne are covered thatare relevant for Fne. This extension is an incrementalprocess. After each extension,evaluation is made on theextended model for behavioral abstraction Fne by usingthe Analysis of Behavioral Patternsheuristic (H2). Thisevaluation indicates that the Bne fails to deliver thebehavioral abstraction Fne because vertical heat transferair-currents are affected. This failure is diagnosed for acause, CS. The cause is a behavioral aspect of Bne that iscausing the loss of vertical heat transfer behavior. The causeis converted as a constraint that is relatable to the behaviorsof the device behavior. It is then posed to the adaptationstrategy of the device model. This returns a new inputbehavior to the new environment. This new behavior, Nb,is added to the behavior Bne. This may result in theelimination of CS from Bne. The extension methodcontinues from the modified Bne. The process repeatsuntil the Analysis of Behavioral Patternsheuristic returnsthat the Bne does not fail to address its behavioralabstraction. The resulting Bne is considered as the finalmodel of the new environment.

The above strategy deals with a situation where the newenvironment can be regionalized into Ske. This may fail intwo situations. Sne is similar to Ske in spatial aspects, but canstill have different structural aspects, e.g., a different medium.In this case, one of the KRITIK’s strategies—substancesubstitution strategy—can be used. In the second situation,Sne is spatially smaller than Ske. In this case, new con-straints are imposed on the behaviors of Bke. This strategyis not considered in this paper, as it is still in development.

4.1.2 Function transferThis method takes the behavioral abstraction of a knownenvironment, and the structural descriptions of known andnew environments. It first identifies the structural aspects, Sf

of Ske, that play a functional role by searching the definitionof Fke. In the behavioral abstraction definition of environ-ment for a refrigerator, as illustrated in Fig. 8, the structuralaspects are source and LCs. That is, the heat source and itslocation are functionally relevant structural aspects for therefrigerator environment. These structural aspects arematched against Sne (Fig. 16).

This results in Sfn¼ { n sources, LCs}. The newbehavioral abstraction definition for the new environmentis constructed, the result of which is illustrated in Fig. 17.Here, LCs, Hs and Ls are all sets.

4.1.3 Behavior regionalization and transferFig. 18 describes a method forregionalizationof Sne andtransferof Bke to Sne, mentioned in Fig. 15. Initially, thespatial dimensions of Ske and Sne are compared. If Ske issmaller in length, height and width, then the regionalizationprocess begins. If the dimensions are equal, then KRITIK’sadaptation strategies are applied. Otherwise, the process halts.

Regionalizationis the method by which a large environ-ment is converted into smaller regions where knownfunctional models of smaller environment are applied.Since functional models such as SBF models allowlocalization and modification, regionalization allowslocalization and modification of a new larger environmentat a number of regions. The Bke behavior model is used toregionalize Sne.

Transferis the method used to generate the models in thenew environment based on the structural similarity with theknown environment. Transfer of generic behavioral patternsfrom one situation to another is not a new concept. TheIDEAL system19,20posits two kinds of behavioral patterns:generic teleological mechanisms (GTMs)31 and genericphysical principles (GPPs).32 IDEAL’s theory of model-based analogy abstracts these behavioral patterns from theSBF model of one physical system, and instantiates them inthe SBF model of another, functionally similar system.

In our case, behavioral transfer is centered around sourcesof Sne, as the sources cause disturbance in the environmentand result in the environment behavior. The Analysis ofStructural Patterns heuristic (H1) suggests not only threesuccessive sources on which Bke can be instantiated, butalso that the first source has to be closest to the right wall ofSne. This selection of sources provides for the widestpossible representative behaviors of Bke. This usefulnessof H1 is based on a representation of the possible behaviorsof sources.

H1 suggests the first source, from the sources of Sne,based on its closeness to the right wall. Here, the transferof behaviors is based on abstraction and instantiation overqualitative regions. When a behavior is transferred to a newenvironment, the qualitative regions in its states are replacedby those of the new environment. Since a qualitative regionconstitutes the component element and the qualitativelocation, these two are replaced in the instantiated behavior.

For some locations in the new environment, even forthose of region R1 (see Fig. 18), the qualitative regionsare not available. For example, the vertical heat exchange


behavior is supposed to exchange heat with the evaporatorat the end of the behavior (see Fig. 11 for behavior). Sincethe height of the air-conditioner environment is much largerthan the height of the refrigerator environment, the para-metric relation (LCe¹ LCs) , (Hs1¹ Htac1) can be vio-lated, resulting in an unpredictable behavioral state. Sincean air-current can reach the ceiling of the refrigerator, pro-vided the required heat gradient is present, an assumption ismade that the air-current reaches the point in Sne, whichcorresponds to the qualitative height of the refrigerator thatsatisfies the parametric relation (LCe¹ LCs) , (Hs1 ¹

Htac1). (The set of such assumptions is passed to the adap-tation strategy of the air-conditioner device.) This is shownas U11 in Fig. 19, which corresponds to an unknown qua-litative region. The unknown qualitative region is aqualitative region where the location may be known butthe behavior of the air-current is not. Beyond this unknownregion, the Bke behavior that has been transferred cannot bepredicted.

A Bke behavior is transferred starting at a knownqualitative region until it reaches an unknown qualitativeregion. The unknown qualitative regions can arise at eitherend of the behavior. In air-current behaviors, illustrated inFigs 10–12, the air-currents exchange heat with source atthe conditional transition (CT) of a state transition. Since thesource is given for each instantiation, the description for thebehavior segment between this state transition and anunknown region and this state transition is not completelyknown. The segments of behavior that are before and after

the state transition with CT are called preceding andsucceeding behavior segments, respectively. Thesepreceding and succeeding behavior segments are called‘tendentious behaviors’. These behaviors have the tendencyto predict the behaviors of air-currents in the region R2,which may decide the behavioral abstraction of Sne. Theuse of tendentious behaviors in generating the behaviors inthe R2 region will be discussed in the next subsection.Fig. 19 gives a complete algorithm for regionalizing Sne,and then transferring Bke to Sne. The result is a newbehavior Bne with three regions instantiated with Bke.

4.1.4 Extending the transferred behaviorsThe major reason for the difficulty in adapting a functionalmodel of an environment to a new environment, withdifferent structure, is due to new qualitative regions arisingin the new environment model. One of the strategiespursued in this research work is to discover such regionsby extendingthe tendentious behaviors in the new environ-ment. The goal of the extension method is not to discover allthe qualitative regions in the behavior of Sne, rather thoseregions which can violate the behavioral abstraction of thenew environment.

The extension method incrementally discovers suchregions. It uses heuristics to discover the new qualitativeregions. Though the extension method can, theoretically,use any heuristics, in this paper we focus upon using theheuristics H1 and H2, as these heuristics have beenimplemented in EKRITIK.

Fig. 15. The overall strategy for adaptation of the environment model.

Fig. 16. The method for generation of function for a new environment.


The extensionmethod uses the Analysis of BehavioralPatterns (H2) heuristic, which states that if a vertical heattransfer behavior (VHTB; Fig. 11) is lost, it can lead to thefailure of the functionality of Sne. The extension methoddiscovers loss of VHTBs by predicting the result of collisionbetween a tendentious VHTB originating in a region with atendentious wall and air-current collision behavior (WACB)originating in a region right next to it. This collision ispossible because WACB, as illustrated in Fig. 12, travelsin the left-up direction. The collision between these two air-currents can result in an air-current that is also left-up. Theregion where there is a collision is called a collision region,e.g., N12 in Fig. 18. These are the new qualitative regions.Two kinds of new qualitative region are possible: theregion where two air-currents collide and the regionwhere an air-current collides with a wall or a source orevaporator. Since the loss of VHTB contributes to thefailure of behavioral abstraction of the new environment,the extension method focusses upon the air-current andair-current collision.

Fig. 20 illustrates the method that extends the Bnebehavior which is generated by the regionalization andtransfer method. This results in a Bne model that does notfail the behavioral abstraction Fne. This method takes asinput the Bke behavior, the Bne behavior and the identityof R1 regions selected by the H1 heuristic. The output of theextension method is the Bne model extended into a new

qualitative region that is not present in Bke. The algorithmillustrated in Fig. 20 includes steps, in addition to those forextension, that show how this Bne output from the extensionmethod could be used for redesigning the Bke model. It alsoincludes steps for coming up with a better model of Bne thateliminates the faults of the Bne model, so that the refinedBne model can satisfy Fne.

Before the algorithm is presented, it is to be noted that thebehaviors of air-currents in the environments have differentroles. Some behaviors enable the behavioral abstraction ofthe environment to be achieved. These are calledenablingbehaviors. VHTB is an example of enabling behavior.Some behaviors distract from achieving the behavioralabstraction and are calleddistracting behaviors. WACB isa distracting behavior. The heuristics H1 and H2 canidentify the roles of the environment behaviors as enablingor distracting. For example, a simple heuristic can decide abehavior as enabling if it has a consistent direction that is amember of {UP, DOWN}. On the other hand, an air-currentbehavior is distracting if it has directions which aremembers of {LEFT-UP, RIGHT-UP, LEFT-DOWN,RIGHT-DOWN} or the behavior has collision nodeswith some of the structural elements of the environmentwhich are not source or device. In the followingdiscussion, we focus upon the example. For instance,instead of using the term enabling behavior, we use theVHTB.

Fig. 17. The function of environment for an air-conditioner.

Fig. 18. The regionalisation, instantiation and extension of Sne.


The R1 region of the Bne set is made into two copies,Regions1 and Regions2, so that from Regions1 the WACBsare selected, and from Regions2, VHTBs are selected. EachWACB selected is tested against each of the VHTBs forcollision.

This collision is tested by extending the tendentiousbehaviors of WACB and VHTB. The tendentious behaviorsare extended by letting the open state (see Fig. 18) begrounded in a structural element that is in the path of theair-current represented by the tendentious behavior. Forexample, the tendentious VHTB (e.g., C31 in Fig. 18),which represents the air rising in the upward direction,can be extended to hit the evaporator in the ceiling (e.g.,E21 in Fig. 18). Similarly, the tendentious WACB(e.g., C12) represents an air-current moving towards theleft wall of the enclosed room. By extension, the air-currentcan hit the left wall of the enclosed room or the ceiling(e.g., E12). These extensions can also be influenced by thetemperature gradients set up by other sources and theevaporator. This latter aspect is not included in the currentimplementation.

After extending the tendentious behaviors, they are testedfor crossover. Crossover of air-currents can be illustrated bytaking two air-currents from Fig. 18. The first air-current(such as C12) moves with direction left-up and towards theleft wall. The second air-current (such as C31) movesupwards towards the ceiling. The first air-current starts ina region right next to the second air-current. Both of theseair-currents can crossover at N12, which represents the new

collision region. This reasoning is used to predict a newqualitative region or collision region. At this region, thebehavior of the resultant air-current, resultant_behavior(e.g., Cac2 in Fig. 18), is predicted. Since there is nobehavior present in Bne to predict the behavior of air-currents at this region, a new method calledcomposingthe extended behaviorsis used. This method learns fromBke as to how two behaviors can be composed. Thismethod is discussed in the next subsection.

The resultant_behavior is tested for reaching the ceiling.Failing this, it is concluded by using the H2 heuristic that thebehavior Bne cannot deliver the behavioral abstraction Fne.The resultant_behavior is diagnosed for the failure toaddress Fne. This cause is then converted into a constraint.This constraint is a behavior that would eliminate thefailure. This constraint is imposed on the redesign of Bndby sending the constraint to its adaptation strategy. Due toredesign of Bke, new inputs are fed into the new environ-ment. These would introduce new behaviors into the newenvironment. These behaviors are added to Bne. Theprocess of extension is then continued. This is repeateduntil there are no more elements in Regions1 and Regions2,signifying that all possible behaviors have been tested forpossible collisions. The resultant Bne behavior delivers thebehavioral abstraction Fne.

4.1.5 Composing the extended behaviorsCollisions of behaviors can result in new behaviors. Anexample of collision behavior for air-current C4 is

Fig. 19. The method for regionalising Sne and instantiating Bke in Sne.


illustrated in Fig. 13, where two air-currents collide witheach other. Another example of collision is the air-currentcolliding with a wall, e.g., air-current C2 in Fig. 12. Duringthe adaptation of an environment model to a model of a newenvironment, air-currents can collide in ways whose repre-sentation is not present explicitly in any of the behaviors ofBke or Bnd. We suggest a method that generalizes therequired principles underlying the behavior for collisionfrom the known Bke model. Our generalization method,generation of composition behavior, has four steps: (i) for-mulate generalization goals from the colliding air-currentsand the structural elements of Sne; (ii) search through Bkefor behavior segments that match with the goals; (iii) arriveat generic physical principles (GPPs) of collision by gener-alizing the matched segments of Bke behaviors; and (iv)finally, instantiate these GPPs into Sne at the collisionregion. The previous subsection described how to arrive ata collision region.

Our method for generation of compositional behaviorsbuilds upon the IDEAL system,19,20,32 which provides acomputational theory of the content, representation,

abstraction, transfer and instantiation of behavioralprinciples in the form of generic teleological principles(GTMs) and GPPs.

Fig. 21 illustrates the method for generalizing GPPs thatare applicable in Bne from Bke. The first step in this methodis to form a goal for generalization. The goal concept to begeneralized is specified in terms of its causal structure, thechanges in property values that have occurred within thecausal structure, and the location where new changesshould occur as a result of the causal relationships. Fig. 22illustrates an example of a goal to find the collision betweentwo air-currents. The causal structure here states how twoair-currents collide at the location N12.

This goal is identified by first recognizing the behavioralstate whose causal consequence is not known and thennoting the causal behavioral states for this state. InFig. 22, the causal consequences of collision state are notknown.

For this goal, there is no corresponding behavior presentin Bne, as the goal requires a causal representation ofcollision of LEFT-UP ascending and UP ascending air-

Fig. 20. The method for extending Bne behaviors.


currents, whereas the behavior in Fig. 13 illustrates thecollision of ascending and descending air-currents. Directsymbol substitution for directions will not give us theresultant air-current. For example, an air-current movingin the UP direction and another one moving in the DOWNdirection can collide to give an air-current that may bemoving in either direction. However, an air-currentmoving LEFT-UP colliding with an air-current movingUP may give only a LEFT-UP moving air-current.Furthermore, other properties like the temperature, heatand the air-current destination temperatures also influencethe direction of the resulting air-current.

The purpose of generalizing GPPs is to generalize theserelationships so that they can be applied to a new situation.For the goals formed for the new collision region, theair-current behaviors are searched for a match. If a matchingbehavior segment is found in the behavior, the behaviorsegment is generalized over the goals and the behaviorin which the behavior segment is found. This generalizationis a GPP. Often, the generalization is made over thebehavioral states and the parametric relation of the statetransition (see Figs 11–14), so that they can be appliedfor the new local situation arising at the new collisionregion.

A new GPP is illustrated in Fig. 23, which is generalizedfrom the goal and the compositional behavior segment ofFig. 13. The causal structure is taken from air-current andair-current collision behavior in Fig. 13. The extent of thecausal structure includes the goals and their consequentstates. However, the states and the state transitions canuse a wide range of properties not included in the behaviorof Fig. 13. For example, the direction of the air-currentdepends not only upon the properties of the air-currentand its location, but also the properties of the sources andthe evaporator. That is, the air-current at point N12 canmove downwards if its temperature is lower than thesource S3. If its temperature is higher than the evaporator,then it will move upwards.

To formulate this generalization, several domainprinciples are used. One of them is the set of Laws of

Geometric Averages, which describes the direction of theresulting air-current. Some of these laws are illustratedbelow.

The direction of the resulting air-current alsodepends upon a function of the temperatures of the air-current, the source and the evaporator, as describedearlier. This function is also part of the domainprinciples. Using these two sets of domain principles,the generalization is made over the known behaviorsegment. The resulting behavior principle is illustrated inFig. 23.

According to the principle generalized, the air-currentproperties such as temperature, direction, heat content andflotation can be worked out using a number of parametricrelations which are explicitly represented in the statetransition. The principle also illustrates that severalresulting air-currents are possible from collision, and eachexhibits different sets of properties. These air-currents maydiffer in their properties depending upon which of thedomain principles can be used. For example, for one air-current, the UPWARD direction principle may be usedand for the other the DOWNWARD direction principle.Which of the domain principles is selected depends uponthe situation into which this behavior principle isinstantiated.

This principle is instantiated in the N12 location or at theend of the {E12 and E31} air-current behaviors of Bne that isbeing developed. This would result in only one air-current,as the Law_of_Geometric_Averages states that the sum ofUP and LEFT-UP air-currents results in only a LEFT-UPair-current.

Fig. 21. The method for generation of composition behaviors.


This ends the stage of arriving at the preliminarymodel of the new environment. Section 7.4 extendsthis EAM model to arrive at a complete design of theair-conditioner model with a fan in the device that caneliminate the cause for the loss of the vertical heat transferbehaviors, namely the wall and air current collisionbehaviors.

5 EKRITIK: IMPLEMENTATION ANDEXPERIMENTATION

The theory described for adaptation of the environmentmodel to a new environment is implemented in a system

called EKRITIK. It is implemented in Allegro CommonLisp running on a Sparc Station 4 using Solaris 6. EKRITIKcompletely implements the ESBF model and the strategiesfor the design of preliminary model of the environment(Section 4.1). EKRITIK makes use of KRITIK fordevice design33 and IDEAL for device redesign.24 BothKRITIK and IDEAL have been tested and their test resultsare reported in the literature. The current EKRITIK imple-mentation focusses on case representation and adaptation,and does not include case retrieval and storage. Since theESBF model is an extension of the SBF model, the retrievaland storage processes can be assumed to be similar.EKRITIK is being extended to implement the environmentmodel evaluation and redesign (see Section 7.4).

Fig. 22. The goal formed for collision region N12 of Fig. 19.

Fig. 23. The GPP of collision of two air-currents.


In EKRITIK, each case, with its functions, behaviors,structure, states and state transitions, is implemented usingthe classes of CLOS of Allegro Common Lisp. Currently, 31generic classes are used to define each case. Each case isdefined using subclasses of these generic classes. Each casein the case-base of EKRITIK is set up before its use.Presently, EKRITIK has a complete implementation of therefrigerator integrated device case described in Section 3.Table 1 gives the various implementation parameters forEKRITIK.

We have tested EAM on the following problems ofengineering inventions. These belong to different domains,require different kinds of adaptations and are of differentcomplexities.

(1) Adaptation of refrigerator to air-conditioner. Thishas been discussed in detail in this paper.

(2) Adaptation of watermill model to windmill model.The ESBF model of a watermill has a devicemodel that describes the conversion of linearmomentum on the fins of an axle to the angularmomentum of the axle. The environment model isthe linear flow of water current in a pipe thatinteracts with the device by transferring the linearmomentum to it. This ESBF model is adapted toan environment where the medium is air and itdoes not have a fixed direction. The new environ-mental modeling suggests that it may not be able todeliver its functionality of transferring the linearmomentum to the fins of the axle due to the changingdirections of the wind currents. This is posed as aconstraint on the redesign of the device model. Theresult is a device where the air-currents of varyingdirection impart momentum to an axle with finsmounted on it. The EAM adds a self-aligningmechanism to the axle to align the fins in a directionperpendicular to the air-current.28

(3) Adaptation of a model of a coffee-maker with hotenvironment to a model of a coffee-maker with acold environment. The coffee-maker makes hotcoffee when the external environment is hot, butfails to deliver hot coffee in a cold environment.The new environment model arrived at by EAMsuggests new thermal paths that make the coffee-maker lose its heat. The strategy suggests severalsolutions, one of which is to provide a heat-resistantcover to block the heat transfer.24,25

(4) Adaptation of a model of a nitric acid cooler with acold water environment to a model of a sulfuric acidcooler with a hot environment. Here, the requiredheat transfer between the sulfuric acid cooler

device and its environment is reduced due to thehigher temperature of the environment. EAMsuggests that the flow rate of the sulfuric acid isreduced to enable better heat exchange.

6 RELATED RESEARCH

Our work is related to and builds on two lines of research inAI: adaptive design and functional modeling. As wedescribed in the introduction, past AI research on adaptivedevice design has focussed on devices that have minimalinteractions with their environments.15,16,18Indeed, we arepresently unaware of any AI work that directly addressesthe task of adapting device designs for operation in newenvironments. In contrast, the issue of device–environmentinteractions is central to the work described here. The EAMstrategy of EKRITIK evolves from the AM strategy used inthe KRITIK system.13,16,17,33

Past AI research on functional modeling has alsoemphasized devices with minimal interaction with theirenvironments.6,7,10,23,34–36In contrast, functional modelingof devices that strongly interact with their environmentslies at the core of our work. The ESBF representation ofEKRITIK evolves from the SBF models used in the KRITIKsystem.10,16,17The SBF representations themselves evolvefrom and unify two earlier device representations: theFunctional Representation scheme of Sembugamoorthyand Chandrasekaran,35 and the Method of Consolidationof Bylander and Chandrasekaran.37 This integration hasgiven the SBF models the properties—localization andmodification—required for adaptive design.

In the work presented in this paper, we extend thesemodels further by introducing the idea of function asresulting from the interactions between the device andenvironments. A considerable amount of research is inprogress investigating different kinds of SBF models. Onesuch work is prototypes,7 where function, behavior andstructure have different variables, and these are mappedonto each other. These mappings are further explained byusing some equations. In our approach, the explanations, i.e.behaviors, integrate all the variables.

The functional model proposed in this paper addressesseveral issues of current interest in functional modeling.The SBF model10 integrates physical principles with theintended function of the device, as in CFRL.6 Our ESBFmodel incorporates physics principles into the integrateddevice and environment. Since our model can incorporateseveral environments, and each environment model canhave a unique set of physical principles, a single integrateddevice and environment model may have several sets ofphysical principles integrated into it. While CFRL allowsdifferent kinds of simulations to be performed on it,our model allows only those simulations that allowexplanation of the function in terms of structuralinteractions.

Our functional model allows representation of multiplemodels as in the multimodeling approach.34 Our model

Table 1. Implementation parameters of EKRITIK

Generic classes 31Classes in refrigerator case . 200No. of parametric values . 30Total length of EKRITIK 6 000 lines (approx.)


represents the functional, behavioral and structural models.While the multimodeling approach is used for diagnosis, ourapproach is used for adaptation in design. Another relatedstream of research is the multilevel flow model (MFM).21

We represent behaviors for interaction with other behaviors.Our modeling is uniform, i.e. we represent the same ontol-ogies for both the device and environment models. We alsouse these models for adaptation in design, unlike diagnosisin MFM.

The devices and environment may constitute differentclasses of structural elements, which show different classesof behaviors. We can consider the device and environmentcombination as a multidomain device.21 In an integrateddevice and environment, the environment has no intendedfunctional abstraction, it has only an ascribed functionalabstraction to support the intended function of the device.This will also eliminate a large number of interactionsthat are possible between the device and its environmentoutside the functionality of the integrated device andenvironment.

Our work addresses some of the important issues thatarise in case-based design. Here, the designer has casesrepresenting past experiences of designing artefacts. Givena new design problem, the designer retrieves the past casesthat match with the new design problem. If this fails, casessimilar to the new design problem are retrieved and areadapted to reduce the differences between the new designproblem and the retrieved case. Two key issues in adapting acase to the new problem are tolocalize aspects of theretrieved case, which may explain the differences betweenthe new design problem and the case, andmodificationofthese aspects so that the differences are reduced.

In KRITIK, 16 the SBF model representing a case isused to localize the functional differences between thenew design problem and the retrieved case. This differ-ence is localized to some behavioral aspects of the case.These localized aspects are modified to generate the adap-tation space, which reduces the functional difference in anumber of ways. In CADET, for localization, a behav-ioral description is elaborated to have new indices, so thatvarious parts of the behavior can be identified andaccessed.15

These systems focus upon the design of devices withminimal environmental interactions, whereas, in thispaper, we focus upon the design of devices with substantialenvironmental interactions. Our adaptation strategy is alsounique in that it is a hierarchical behavioral adaptation: themodels for device and environments can be described atdifferent grain levels and their adaptations take place attheir grain levels. In our methodology, the knowledge thatis transferred from past cases to a new design problem,in order to accomplish adaptation, is an abstraction andinstantiation of behavioral models in the past cases. Ourmodels express an important intuition that when a designerdesigns a device, he/she understands how the device givesrise to its function. This understanding, though specific to adevice, can express several principles that can be used inother devices.

7 DISCUSSION

The major objective of this paper has been to extend thefunctional models of the physical systems to include theirenvironments. Functional modeling of physical systems islimited if the interactions between the devices and theirenvironments are not considered. Our second objectivehas been to understand the influence of functional modelingof this interaction on the design of engineering devicesthrough adaptation.

The functional model we have developed, as an extensionof the SBF models, not only explicitly models the environ-ment, but also models the interactions between the deviceand its environment for the functioning of the device. Inother words, the function of a device can be defined as thedesigner’s goal for the interaction between the device and itsenvironment.

The ESBF model is adaptable by having two properties:functional decomposition between the device and itsenvironment, and a causal representation within the modelsof the device and its environment. These enable easylocalization of functional differences onto local modelaspects. The causal models enable generation of adaptationspaces by modification of the local aspects within theenvironment and the device models.

The device and environment models are inherentlydifferent, even though they share the same SBF model.The structural differences can influence their behaviorsdifferently. In devices, structural variations can causelocalizable behaviors. In environments, global changes inoutput behavior often result. This has lead to very differentadaptation strategies for both. In device model adaptation,the local model aspects are modified by using anequivalence class relation. In environment modeladaptation, local aspects are generalized from a knownenvironment model and combinations of such local aspectsare discovered to achieve the desired output behavior.

We have used adaptation for the design of devices, whichmay mean that either the known device or its environmentmodel need to be modified to achieve a new functionality.Our work addresses some of the important issues that arisein case-based design.13,14,38 The changes in device andenvironment models due to adaptation are interdependent,as together they support the functionality of the device. Inorder to support this interaction, even during adaptation, wehave introduced a constraint exchange and satisfactionmechanism.

7.1 Limits on interaction

Devices differ in the amount of interaction they need tohave with their environments in order to support theirfunctionality. At the low end of the spectrum, we have anelectric buzzer circuit, where the interactions are only at theinput and output of the circuit for a short period of time. Atthe other end of the spectrum, as in an airplane, there is acontinuous interaction at a large number of points on thesurface of the plane. A large number of devices, such as a


refrigerator and a watermill, fall between these twoextremes. At the upper end of interaction spectrum, thesemodels may not be very useful for adaptation. In themiddle of the interaction spectrum, it is possible toform models with high levels of abstraction that can beused in adaptation. ESBF models address this class ofinteractions.

7.2 Limits on adaptation

The adaptation and constraint satisfaction stages ofadaptation strategy should yield new integrated device andenvironment models that are adaptable. This condition maybe violated if the constraints already satisfied by the knownintegrated device and environment models are violated.Since each of the known device and environment modelscould have some constraints incorporated into them, anynew constraints that need to be satisfied during a new designshould not be in conflict with them. The current strategiescannot resolve such constraint conflicts.

Another problem arises due to adaptation of functionaldecomposition to new device design. This functionaldecomposition may not be applicable to new designs. Forexample, to design, from a refrigerator, a nitric acid cooler,one is forced to view the water flowing inside a chamber asthe environment. The adaptation strategies discussed in thispaper are successful where functional decompositionsbetween the device and its environment are similar.

7.3 Scalability of ESBF models

The ESBF model is inherently modular: the models ofdevice and environment are functionally independent. How-ever, they are based on the same SBF model. Due to this, itis possible to scale up the computational models. Scaling upmay be required due to either large devices or largeenvironments, or large functional differences between theknown integrated device and environment and the new ones.Both complex devices and complex environments can beconsidered as made up of several devices and environments,as illustrated in Fig. 24. For example, a refrigerator can havea very complex environment. This can be modularlyconsidered as follows: a refrigerator is inside a room andthe room is inside a house, etc. Large functional differencesare harder to address. In the current computational model,similarities between functions of known and new integrateddevices and environments can result in successful designs.

7.4 Extensions to EAM: evaluation and redesign

Section 4.1 presents an aspect of EAM which arrives at thepreliminary design of a model of the new environment. Thisaspect has been implemented in EKRITIK. However, thismodel does not arrive at the completed design of theintegrated device and environment. In this subsection, wepresent an extension to EAM that provides a completemodel of the integrated device and environment.Many parts of this extension have been developed and

implemented elsewhere. For example, the device redesignhas been dealt with in Prabhakar and Goel25 and Bhattaet al.24 Furthermore, the principles necessary to developthis extension are already present in preliminary design(Section 4.1).

The strategy for adapting the model of an environment toa new environment, described in Figs 15, and 20, arrives at apreliminary model Bne of the new environment, thenevaluates it and finally modifies it to satisfy its function.In the earlier section, we discussed various methodsinvolved in developing the preliminary model of the newenvironment. This development was based upon theheuristic search aiding another search in the adaptationspace of causal models. In this section, we discuss methodsto evaluate the Bne model, its redesign and the redesign ofthe device model Bnd.

7.4.1 Evaluation and diagnosis of BneThe first step is to evaluate whether the Bne developedwill be able to deliver the function Fne. As we mentionedearlier, it is too complex to completely simulate anenvironment, hence a heuristic method is used. The heuristicused is the Analysis of Behavioral Patterns (H2). Thisstates that any loss of vertical heat transfer air-currentbehaviors can lead to failure of the functionality of theenvironment.

For Bne, developed for an air-conditioner, the H2heuristic suggests that it will fail to deliver thefunctionality because one of the representative verticalheat transfer behaviors {C31, E31} can be lost due to itscollision with the behavior for the air-current E12 (see Fig.18). This is because the resultant behavior found for thecollision of these two air-currents can have the directionLEFT-UP.

The next stage of evaluation is diagnosing the causeof the failure of the function. The Bne behaviormodel, with the instantiated behavior principle of Fig. 23,is inspected to identify a cause. The following arethe possible causes for failure of Bne behavior of anair-conditioner.

7.4.2 Constraint formation and postingSince the designer is not supposed to change the environ-ment, the failure of the environment to deliver its function-ality is translated as failure of the device to supply adequateinputs to its environment. In other words, it may be seen asthe device providing an inappropriate kind of interactionwith its environment. The device should allow for a differ-ent kind of interaction with the environment, which mayinclude new properties, or the device attempts to eliminatefailure-causing behaviors of the environment. It is the latter


strategy we adapt in this paper, as the former requiresremodeling the environment for new properties.

The major issue in constraint formation is to link theenvironment and device models, as they can be based ondifferent properties. For example, the device does not have acomponent called air-current, whereas the environment has.The constraint formation strategy translates the cause of theenvironment model failure to the constraints on the device.The strategy uses the domain knowledge of both thedevice and the environment. The following is a list ofconstraints formed for the causes identified earlier for theair-conditioner.

These constraints are passed to the adaptation strategy ofthe device so that Bnd can be redesigned to satisfy theconstraints.

7.4.3 Redesigning device behavior modelBefore we begin with a description of the redesign processof Bnd, we briefly mention how the device model is adaptedfrom the known device model. This may involve applyingthe adaptation strategies of KRITIK. Since the refrigeratorand air-conditioner share the same function, the adaptationdoes not involve a change to the behavior model presentedin Figs 6 and 7.

The constraints are passed to the adaptation strategy forBnd and are incorporated into Bnd design by using themethod described in Fig. 25. As a first step, the behavioralsegment of Bnd that interacts with the environment isidentified. In Figs 6 and 7, this corresponds to the statetransitions, where interaction with the air-currents isindicated. Since we already have a model of how the inter-action occurs between the device and its environment, it ispossible to localize to those aspects of interaction where theconstraints are applicable. For example, the constraint (Cn1above) that requires the device to generate an air-currentthat can eliminate collisions in the environment can beapplied on the state transition of Bnd, where the interactionwith the environment occurs. Since this transition fails todeliver the constraint, this is formulated as a new functionwhich the state transition should deliver.

The design_case_base is searched for the new function

and the retrieved behavior is composed of the statetransition of Bnd, where the interaction with the environ-ment occurs. A possible behavior that can address the Cn1constraint can be a ‘pump’ that pumps the air-currents in aparticular direction. The Bnd, composed with the retrievedbehavior, is further evaluated to satisfy the constraint that isbeing addressed and the function Fnd. If they are satisfied,then control is passed back to redesign of Sne, else the Bndis diagnosed for a cause. This is repaired. A more detaileddescription of this method and an example can be found inPrabhakar and Goel25 and Bhattaet al.24

7.4.4 Redesign of environment modelThe new Bnd adds new behaviors to the Bne. For example,the pump behavior added to Bnd sets up new behaviors tothe environment in the direction RIGHT-DOWN. Thesebehaviors are extended and composed with the existingbehaviors of the Bne using the principles of instantiation,extension and composition discussed in the earlier section.The resulting behavior gives a revised model for Bne. Thisbehavior model is evaluated for the functionality and thecycle repeats, as illustrated in Fig. 25.

8 CONCLUSIONS

Computational methods based on functional models ofdevices have proved to be very useful for automatingaspects of adaptive design. However, past AI research onfunctional modeling and adaptive design has been limited todevices with minimal interactions with their environments.We have developed a functional representation scheme,called the Environmentally-bound Structure–Behavior–Function (ESBF) model, to represent and organizeknowledge of the functioning of a device, including therole of its environmental interactions. We have alsodeveloped a processing strategy called Environmentally-driven Adaptive Modeling (EAM) for modifying a knowndesign for operation in a new environment, and thuscarrying out new functions that arise due to new device–environment interactions. ESBF and EAM arise from ananalysis of three case studies of technological innovation:the invention of the windmill, the air-conditioner and theautomatic coffee-maker. EKRITIK implements ESBF andEAM for the example of the evolution of the air-conditionerfrom the refrigerator.

A key result of EKRITIK is a recharacterization of thefunction of a device. Instead of adopting the traditional viewof a device function as an abstraction of the internal causalbehavior of the device, ESBF views it as an abstraction ofthe interaction of the device with its external environment.This new view forms the foundation of EAM.

ESBF models have several other desirable features:functional closure of the device behavior, causal closureover device structure, equivalence of causal events overdevice and environmental behaviors, and representation ofacausal interactions between the device and the environ-ment. Spatial interactions between air-currents in a room

Fig. 24. The scalable property of environmentally-bound SBFmodels.


provide an example of the latter. ESBF models provideconceptual primitives for representing such spatial inter-actions. These features enable EAM to localize functionalaspects of the design onto its structural elements and isolatemodifications to the structure to obtain new behaviors andfunctions.

ACKNOWLEDGEMENTS

The initial conceptual development of the work was donewhen S. Prabhakar was visiting A. K. Goel’s group atGeorgia Institute of Technology on sabbatical leave fromthe University of Technology in 1995. This work benefittedfrom many discussions with members of A. K. Goel’sDesign Intelligence group and especially from discussionswith and contributions from Sambasiva Bhatta. We thankthe Editors of this special issue—Drs Luca Chittaro andAmruth Kumar—for their support and encouragement,and the anonymous reviewers for their constructive feed-back. During the writing of this article, A. K. Goel wassupported in part by design-related research grants fromNSF (DMI-94-20405) and DARPA (F33615-93-1-1338).

REFERENCES

1. McDermott, J. R1: a rule-based configurer of computersystems.Artificial Intelligence, 1982,19, 39–88.

2. Brown, D. C. & Chandrasekaran, B.,Design ProblemSolving: Knowledge Structures and Control Strategies.Morgan Kaufmann, San Mateo, CA, 1989.

3. Mittal, S., Dym, C. and Morjaria, M. PRIDE: an expertsystem for the design of paper handling systems.IEEEComputer, 1986,19(7), 102–114.

4. Steinberg, L. I., Design as refinement plus constraintpropagation: the VEXED experience. InProceedings of the6th National Conference on Artificial Intelligence, 1987, pp.830–835.

5. Marcus, S., Stout, J. and McDermott, J. VT: an expertelevator designer that uses knowledge-based backtracking.AI Magazine, 1987,8(4), 41–58.

6. Vescovi, M., Iwasaki, Y. & Fikes, R., CFRL: a language forspecifying the causal functionality of engineering devices. InProceedings of AAAI-93, 1993.

7. Gero, J. Design prototypes: a knowledge representationschema for design.AI Magazine, 1990,11(4), 26–36.

8. Umeda, Y., Ishii, M., Yoshioka, M. & Tomiyama, T.,Experimental use and extension of the FBS modeller. InRepresenting and Reasoning about Device FunctionWorkshop Notes, AAAI-94, Seattle, WA, 1994.

9. Prabhakar, S. & Goel, A. (eds),Workshop Notes onFunctional Modelling in Design, Artificial Intelligence inDesign Conference ’96, Stanford University, USA, 1996.

10. Goel, A., Model revision: a theory of incremental modellearning. InProceedings of the 8th International Conferenceon Machine Learning, Chicago, June 1991. MorganKaufmann, Los Altos, CA, pp. 605–609.

11. Balazs, M. E. & Brown, D. C., Function in designpresentations. InWorkshop Notes on Functional Modellingin Design, Artificial Intelligence in Design Conference ’96,ed. S. Prabhakar & A. Goel. Stanford University, USA, 1996.

12. Chandrasekaran, B. & Josephson, J. R., A definition ofdevice function for design activity. InWorkshop Notes onFunctional Modelling in Design, Artificial Intelligence inDesign Conference ’96, ed. S. Prabhakar & A. Goel. StanfordUniversity, USA, 1996.

13. Goel, A. & Chandrasekaran, B., Case-based design: a taskanalysis. InArtificial Intelligence Approaches to EngineeringDesign, Volume II: Innovative Design, ed. C. Tong & D.Sriram. Academic, San Diego, 1992, pp. 165–184.

14. Navinchandra, D.,Exploration and Innovation in Design:Towards a Computational Model. Springer, New York, 1991.

15. Navinchandra, D., Sycara, K. and Narasimhan, S. Atransformational approach to case-based synthesis.AIEDAM, 1991,5(1), 31–45.

Fig. 25. A method for redesigning the device model to satisfy the environment constraints.


16. Goel, A., A model-based approach to case adaptation. inProceedings of the 13th Annual Conference of the CognitiveScience Society, Chicago, August 1991. Lawrence Erlbaum,Hillsdale, NJ, 1991, pp. 143–148.

17. Goel, A. & Chandrasekaran, B., Functional representation ofdesigns and redesign problem solving. InProceedings of the11th International Joint Conference on Artificial Intelligence(IJCAI-89), Detroit, MI, August 1989. Morgan Kaufmann,Los Altos, CA, 1989, pp. 1388–1394.

18. Sun, A. & Faltings, B., Supporting creative mechanicaldesign. InProceedings of Artificial Intelligence in DesignConference 1994, ed. J. Gero and F. Sudweeks. Kluwer,The Netherlands, 1994.

19. Bhatta, S. & Goel, A., Learning generic mechanisms fromexperiences for analogical reasoning. InProceedings of the15th Annual Conference of the Cognitive Science Society,Boulder, CO, 1993, pp. 237–242.

20. Bhatta, S. & Goel, A., From design cases to genericmechanisms.Artificial Intelligence in Engineering Design,Analysis and Manufacturing, Special Issue on MachineLearning, 1996,10, 131–136.

21. Larsson, J. E. Diagnosis based on explicit means-end models.Artificial Intelligence, 1996,80, 29–93.

22. Chittaro, L., Functional diagnosis and prescription of mea-surements using efforts and flow variables.IEE ControlTheory and Applications, 1995,142, 420–432.

23. Kumar, A. and Upadhyaya, S. Function-based candidatediscrimination during model-based diagnosis.AppliedArtificial Intelligence, 1995,9(1), 65–80.

24. Bhatta, S., Goel, A. & Prabhakar, S., Innovation in analogicaldesign: a model-based approach. InArtificial Intelligence inDesign ’94, ed. J. S. Gero & F. Sudweeks. Kluwer, TheNetherlands, 1994, pp. 57–74.

25. Prabhakar, S. & Goel, A. K., Performance driven creativity indesign: constraint discovery, model revision, and casecomposition. In Second International Round-TableConference on Computational Models of CreativeDesign, ed. J. S. Gero & F. Sudweeks, December 1992,pp. 101–127.

26. Prabhakar, S. & Goel, A., Learning about novel operatingenvironments: designing by adaptive modelling. InArtificialIntelligence in Design, Analysis and Manufacturing, Vol. 10.Cambridge University Press, 1996, pp. 151–156.

27. Prabhakar, S. & Goel, A., Extending functional models fornon-local adaptations in engineering device design. InProceedings of the 9th Florida Artificial Intelligence

Research Symposium, Key West, FL, May 1996, pp. 369–374.

28. Prabhakar, S., Goel, A. & Bhatta, S., Adaptive modelling inthe design of interactive devices. InThird InternationalRound-Table Conference on Computational Models ofCreative Design, ed. J. S. Gero and F. Sudweeks, December1995, pp. 267–302.

29. Langley, B. C.,Refrigeration and Air Conditioning. Reston,Reston, VA, 1978.

30. Trott, A. R.,Refrigeration and Air-Conditioning, 2nd edn.Butterworth, London, 1989.

31. Stroulia, E. & Goel, A., Generic teleological mechanisms andtheir use in case-adaptation. InProceedigns of the 14thAnnual Conference of the Cognitive Science Society,Bloomington, IN, 1992, pp. 319–324.

32. Bhatta, S. & Goel, A., Model-based discovery of physicalprinciples from design experiences.Artificial Intelligencefor Engineering Design, Analysis and Manufacturing,Special Issue on Machine Learning in Design, 1994, 8(2),113–123.

33. Goel, A. & Prabhakar, S., A control architecture for redesignand design verification. InProceedings of the 1994 SecondAustralian and New Zealand Conference on IntelligentInformation Systems, Brisbane, Australia, 1994, pp. 377–381.

34. Chittaro, L., Guida, G., Tasso, C. and Toppano, E. Functionaland teleological knowledge in the multimodelling approachfor reasoning about physical systems: a case study.IEEETransactions on Systems, Man and Cybernetics, 1993,23(6), 1718–1751.

35. Sembugamoorthy, V. & Chandrasekaran, B., Functionalrepresentation of devices and compilation of diagnostic pro-blem solving systems. InExperience, Memory andReasoning, ed. J. Kolodner and C. Riesbeck. LawrenceErlbaum, Hillsdale, NJ, 1986.

36. Umeda, Y., Takeda, T., Tomiyama, T. & Yoshikawa, H.,Function behavior and structure. InProceedings of 5thIEEE Conference on Artificial Intelligence Applications,Vol. 1. IEEE Press, 1990, pp. 177–193.

37. Bylander, T. & Chandrasekaran, B., Understanding behaviorusing consolidation. InProceedings of the 9th InternationalJoint Conference on Artificial Intelligence, 1985, pp. 450–454.

38. Hinrichs, T. & Kolodner, J., The roles of adaptation in case-based design. InProceedings of the AAAI-91, AAAI Press/MIT Press, Cambridge, MA, 1991.
