Tw complemen tary patterns to build ultiexpert systems Philippe Lalanda ThomsonCSF CorporateResearc Laboratory Phone  Email lalandathomsonlcr
108K - views

Tw complemen tary patterns to build ultiexpert systems Philippe Lalanda ThomsonCSF CorporateResearc Laboratory Phone Email lalandathomsonlcr

fr Domaine de Corbeville F91404Orsa rance Abstract The purp ose of this pap er is to presen tt o complemen tary patterns to build m ultiexp ert systems that is systems that need to in tegrate large heterogeneous sp ecialized mo dules and implemen com

Download Pdf

Tw complemen tary patterns to build ultiexpert systems Philippe Lalanda ThomsonCSF CorporateResearc Laboratory Phone Email lalandathomsonlcr




Download Pdf - The PPT/PDF document "Tw complemen tary patterns to build ulti..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.



Presentation on theme: "Tw complemen tary patterns to build ultiexpert systems Philippe Lalanda ThomsonCSF CorporateResearc Laboratory Phone Email lalandathomsonlcr"— Presentation transcript:


Page 1
Tw complemen tary patterns to build ulti-expert systems Philippe Lalanda Thomson-CSF CorporateResearc Laboratory Phone: 33169339290 E-mail: lalanda@thomson-lcr.fr Domaine de Corbeville F-91404Orsa rance Abstract The purp ose of this pap er is to presen tt o complemen tary patterns to build m ulti-exp ert systems, that is systems that need to in tegrate large, heterogeneous sp ecialized mo dules, and implemen complex, non deterministic con trol strategies. The rst pattern summarizes the blac kb oard pattern presen ted in [1 ]. The second pattern builds on the mo del of dynamic

con trol where a blac kb oard system has a rep ertoire of indep enden t domain and con trol kno wledge sources, con trol plan expressing the system's desirable b eha vior, and meta-con troller that ho oses at eac h p oin in time the curren tly enabled kno wledge source that b est matc hes the con trol plan. A system implemen ting this arc hitectural pattern is able to dynamically c hange its con trol strategies as resp onse to unpredictable ev en ts. Area: Con trol in complex, dynamic systems Keyw ords: Arc hitectural pattern, blac kb oard systems, dynamic con trol In troduction The blac kb

oard arc hitectural pattern pro vides a computational framew ork for the design and implemen tation of systems that need to in tegrate large and div erse sp ecialized mo dules, and implemen t complex, non deterministic con trol strategies. This pattern, describ ed in [1], de nes sev eral comp onen ts, of whic h the con trol comp onen tis the most imp ortan t. Sev eral mo dels ha e b een prop osed to supp ort the de nition of this comp onen t that is k ey for the system's global prop erties. The purp ose of this pap er is to presen tt o complemen tary patterns. The rst pattern summarizes the

blac kb oard pattern presen ted in [1 ]. The second pattern is concerned with the blac kb oard pattern's con trol comp onen t. It presen ts an approac building on mo del of dynamic con trol that allo ws con trol strategies to be hanged dynamically in resp onse to unpredictable in ternal and external ev en ts. Blac kboard arc hitectural pattern Name Blac kb oard Example Consider an autonomous rob ot running in an oce en vironmen t. It p erforms tasks lik e fetc hing ob jects, sending messages, guiding visitors, and nigh t-time surv eillance. In so doing, it in teracts with uman b eings and

other dynamic en tities and acquires p oten tially useful kno wledge of its en vironmen t. The rob ot has to con tin uously p erform three distinct functions: It p erceiv es its dynamic en vironmen t in order to mo e and to get orders from h uman b eings,
Page 2
It reasons in order to in terpret the p erceiv ed data, solv e problems and determine actions to b e triggered, It acts on the en vironmen t. Di eren functions to be p erformed require v arious exp ertise. Mo dules sp ecialized in arious domains suc as data in terpretation or route planning ha e to b e dynamically com bined

so that the rob ot can ac hiev e its task. Con text Soft are systems that need to in tegrate large and div erse sp ecialized mo dules, and implemen complex, non deterministic con trol strategies. Problem Classical kno wledge-based systems lik e exp ert systems tac kle problems that do not ha e a feasible, deterministic solution. Ho ev er, these systems are not able to deal with complex applications suc h as signal in terpretation, pro cess con trol, or autonomous mobile rob ots con trol. The main reason is that suc applications ha ery stringen t requiremen ts including: The problem under

consideration spans sev eral elds of exp ertise In termediate solutions require di eren t represen tations and paradigms The con trol strategy is complex and cannot b e determined statically Solution The blac kb oard paradigm de nes heterogeneous problem solving represen tations as indep enden t mo dules called kno wledge sources. Kno wledge sources can b e seen as sp ecialists in sub- elds of the global application and are only able to solv e sub-problems. They read and write relev an t data in a blackb ar whic h is a structured global memory where a solution to the problem under

consideration is incremen tally constructed. When a kno wledge source pro duces a signi can tc hange in the blac kb oard, it generates an event Blackboard Control KS-1 KS-2 KS-n . . . . Figure 1: The blac kb oard mo del Eac h kno wledge source has a set of triggering conditions that can b e satis ed b y particular kinds of ev en ts, that is global c hanges in the blac kb oard resulting from external inputs or previously executed kno wledge sources. When an ev en t satis es a kno wledge source's triggering conditions, the kno wledge source is enabled and its parameters b ound to v ariable v

alues from the triggering situation. A giv en kno wledge source will b e enabled, and therefore executable, whenev er ev en ts satisfying its triggering conditions o ccur, regardless of its relativ e utilit yinac hieving the curren t goals. eac p oin in time, man comp eting kno wledge sources ma be enabled. It is the purp ose of the con trol comp onen t to select the b est one for immediate execution. The problem solving pro cess is opp ortunistic in that
Page 3
sense that the activ ations of kno wledge sources are not sc heduled in adv ance, but determined at ev ery con trol cycle

dep ending on the curren t situation. In the basic blac kb oard mo del, con trol strategies are xed. The general organization of blac kb oard system is illustrated gure 1. Dash lines represen con trol ux whereas solid lines stand for read/write accesses. Robot example: Soft are systems con troling an autonomous rob ot ha e to p erform cognitiv e and ph ysical tasks, including: Na vigation En vironmen t learning Reasoning (ab out the goals that ha eto be ac hiev ed) Destination planning Route planning Execution monitoring These arious tasks can be ac hiev ed one or sev eral sp ecialized kno

wledge sources. or example, system usually includes di eren tna vigation routines making use of v arious com binations of sensors, v arious strategies wal l fol lowing is an example of strategy), and v arious prop erties suc h as safet y or sp eed. The blac kb oard con tains global information suc h as the curren t goals, destination, route and constrain ts. It also includes a map whic h stores static information ab out the en vironmen t and dynamic information related to the rob ot kno wledge and previous activities (lev el of kno wledge of a giv en zone, routes already tak en, etc... ). In

the basic approac h, the con trol comp onen is complex sc heduler implemen ting xed strategy eac con trol cycle, that is after eac h kno wledge source execution, it computes a priorit y (a rating) for ev ery enabled kno wledge source. The rating is computed using set of heuristics Examples of heuristics are navigation outine always gets etter ating than le arning outine or when \wal l fol lowing" navigation outine is exe cutable, it r eveives the highest r ating among navigation r outines Structure The blac kb oard mo del de nes three main comp onen ts: The blac kb oard is a structured global

memory con taining ob jects from the solution space. These ob jects, also called blac kb oard no des (or h yp otheses), are hierarc hically organized in to lev els of analysis and can b e link ed with eac h other. Kno wledge sources can b e seen as highly sp ecialized mo dules with their o wn represen tation. They are c har- acterized b y a set of triggering conditions and an executable co de that retriev es data from the blac kb oard and adds its con tribution to the problem solving pro cess. The con trol comp onen selects, con gures and executes kno wledge sources. Determination of

executable kno wledge sources is based on the state of the problem solving pro cess as expressed in the blac kb oard. Structural relationships b et een these comp onen ts are summarized in Figure 2. Implemen tation The rst step when designing blac kb oard-based system is to de ne the solution space, that is the arious in termediate solutions and their represen tation. This activit y leads to the de nition of the blac kb oard structure. It is then necessary to iden tify the kno wledge sources that can pro vide these solutions. Ob viously these activities are v ery related and in uence eac h

other. Kno wledge sources ma y corresp ond to existing mo dules. In that case, mo dules ha eto be put in the form of kno wledge sources. That means that triggering conditions ha e to b e added, input v ariables ha e to b e link ed to blac kb oard data so that they can b e b ound at run time, and results ha e to put in the blac kb oard. This last adaptation is generally made b y simple co de wrapping.
Page 4
update() blackboardNodes access() Blackboard 1+ Knowledge Sources updateBlackboard() execCondition() execAction() operates on 1+ Control selectKS() configureKS() executeKS()

activates reads Figure 2: System's structure (adapted from [1]) The next step is to sp ecify the con trol comp onen t. It generally tak es the form of a complex sc heduler that mak es use of a set of domain-sp eci c heuristics to rate the relev ance of executable kno wledge sources. uning heuristics is a v ery time-consuming activit It is ho ev er a crucial setting that determines the system's suitabilit Kno wn Uses Blac kb oard-based systems ha e b een used in n umerous domains, including: Sp eec h recognition [2] ehicle iden ti cation and trac king [3] Iden ti cation of the structure of

protein molecules [4] Sonar signals in terpretation [5] Consequences The blac kb oard mo del pro vides e ectiv e solution to the design and implemen tation of complex systems where heterogeneous mo dules ha e to b e dynamically com bined to solv e a problem. It also ensures v ery imp ortan t non functional prop erties suc h as: Reusabilit Kno wledge sources are indep enden t sp ecialists that can b e reused in di eren t pro jects. Reuse is made easier b y the fact that there is no direct comm unication b et een kno wledge sources. Changeabilit y and main tainabilit High lev el of mo

dularization and clear separation b et een con trol and domain ( i.e. the kno wledge sources) mak es the main tenance phase easier. Robustness. The mo del naturally leads to the de nition of alternativ e kno wledge sources to solv e a giv en sub-problem. The blac kb oard arc hitectural pattern has also some w eaknesses. The main one is certainly the dicult y of de ning a good con trol strategy Con trol is based on heuristics tuning that can only b e obtained through exp erimen ts. Credits The blac kb oard mo del has b een de ned y the mem b ers of the HEARSA Y-I I pro ject [2 ], and rst

applied on sp eec h recognition. Comprehensiv e presen tation of the blac kb oard paradigm and existing systems can b e found in [5 ] [6]. The rst pattern description of the blac kb oard mo del has b een presen ted in [1 ].
Page 5
Blac kboard based con trol pattern Name Blac kb oard-based con trol Example Consider the autonomous rob ot running in an oce en vironmen in tro duced in the example section of the blac kb oard pattern. Some autonomous rob ots need to in terlea e activities dep ending on the external situation and their curren t state in terms of goals, resources,

information a ailable. or example, supp ose a rob ot is on its w y to fetc h a book with no hard deadline. The rob ot ma y decide to follo w a route it has not tak en for a while in order to up date its kno wledge ab out it. Supp ose no w that it receiv es a request to go to a giv en ro om as so on as p ossible. It will then drop its information collecting task and tak e the quic er and safer route to reac h the ro om. This has an e ect on the na vigation routines it selects to mo e, the data it p erceiv es, and the reasoning tasks it p erforms. More generally , it in uences its whole con trol

strategy The oce rob ot illustrates a class of systems that ha e to ensure the co op eration of a set of comp onen ts in order to solv ev arious problems and that m ust adapt dynamically their con trol strategies. Con text system that in tegrates heterogeneous sp ecialized mo dules and that needs to dynamically hange its con trol strategies as a resp onse to unpredictable ev en ts (in ternal or external). Problem The blac kb oard mo del is a problem solving mo del that separates domain kno wledge in to heterogeneous mo dules called kno wledge sources. A con trol comp onen t pro vides mec

hanisms to activ ate kno wledge sources at the righ time in order to ensure e ectiv e co op eration b et een them. Selection of the b est enabled kno wledge source is based on sp eci c criteria and generally demands v ery precise tuning. In most blac kb oard systems, these criteria cannot b e c hanged dynamically Changing these criteria at run-time is ho ev er necessary in man applications lik rob ot motion in uncertain en vironmen t, aircraft pilot advising, pro cess con trol, or in tensiv care monitoring. These applications require highly adaptiv e systems that are able to adapt their

meta-con trol strategies to dynamic con guration of demands, opp ortunities, and resources for b eha vior. Solution The prop osed solution builds on the mo del of dynamic con trol where a blac kb oard system has (a) a rep ertoire of indep enden t domain and con trol kno wledge sources that are describ ed in term of their resource requiremen ts and result prop erties; (b) a con trol plan expressing its desirable b eha vior; (c) a meta-con troller that c ho oses at eac p oin t in time the curren tly enabled kno wledge source, domain or con trol, that b est matc hes the curren t con trol plan. In

this approac h, criteria guiding the selection of executable kno wledge sources are encapsulated in con trol plan. A con trol plan describ es the system's in tended b eha vior as a temp oral graph of plan steps, eac h of whic comprises a start condition, a stop condition, and an in tended activit y in the form of a tuple (task, parameters, constrain ts). Con trol plans do not refer explicitly to an y particular kno wledge source in the system's rep ertoire. They only describ e in tended b eha viors in terms of the desired task, parameter v alues, and constrain ts. Th us, at eac p oin in time,

the system has plan of in tended action, whic in ten tionally describ es an equiv alence class of desirable b eha viors and in whic h curren tly enabled sp eci c kno wledge sources ma yha e graded degrees of mem b erships. Con trol kno wledge sources are describ ed and managed in the same as domain kno wledge sources. Eac kno wledge source has an in terface that describ es the kinds of ev en ts that enable it, but also the v ariables to b e
Page 6
Goal: plan-destination-sequence Constraints: (fast) Variables: (place-sequence, path) Goal: plan-routes Constraints: (fast) Constraints:

(fast) Goal: navigate Variables: (path) Variables: (place-set, place-sequence) Figure 3: Example of con trol plan b ound in its enabling con text, the task it p erforms, the yp e of metho d it applies, its required resources (e.g, computation, p erceptual data), its execution prop erties (e.g, sp eed, complexit completeness), and its results prop erties. Domain and con trol kno wledge sources do not access and up date the same kno wledge base. Domain kno wl- edge sources are concerned with the solving of a particular problem. Con trol kno wledge sources deal with the managemen t of the system

con trol plan. They can replace the curren t plan, p ostp one it, re ne it, etc... The meta-con troller attempts to follo the curren con trol plan executing the most appropriate enabled kno wledge sources. Sp eci cally , at eac h p oin t in time, the meta-con troller con gures and executes the enabled kno wledge source that: (a) is capable of p erforming the curren tly planned task with the sp eci ed parameteriza- tion; and (b) has a description that matc hes the sp eci ed constrain ts b etter than an y other enabled kno wledge source that also satisfy (a). If the selected enabled kno wledge

source is a con trol kno wledge source, the con trol strategy is up dated. Otherwise, a domain kno wledge source is executed in order to con tribute to the problem solving pro cess. Robot example: Figure 3 pro vides a simple example of con trol plan. This plan corresp onds to a situation where the rob ot's high lev el goal is to visit a set of places as so on as p ossible. The successiv e sub-goals that go ern the rob ot's actions are to decide on a sequence of places to visit, to compute the b est route, and to na vigate with a constrain t of rapidit Note that, for the sak e of clarit ,w e

did not include the actions to b e undertak en b y the rob ot at the visited places in the con trol plan (this requires to add a lo op na vigate / c hec k in the con trol plan). Suc plan allo ws the meta-con troller to select the b est enabled kno wledge source at eac p oin in time. It is set b y a con trol kno wledge source that made it up or instanciated it from a library of skeletal plans When the curren t goal is ac hiev ed (here, all the sc heduled places ha e b een visited) or new conditions app ear (new goal, imp ortan tc hange in the en vironmen t, etc... ), another con trol kno wledge

source can b e enabled in order to p ost a more appropriate con trol plan. or example, in a less stressed situation, a new plan could insist on the imp ortance of learning the en vironmen while mo ving. This w ould result in the selection of di eren t paths, and di eren tna vigation routines. Structure The system is divided in to v e ma jor comp onen ts: The blac kb oard is the shared data structure where solutions are built The con trol plan encapsulates meta-con trol information necessary to run the system under the form of a temp oral graph of plan steps. It is accessed and up dated b y con

trol kno wledge sources. Domain kno wledge sources are concerned with the solving of domain-sp eci c problems. Con trol kno wledge sources adapt the curren t con trol plan to the curren t situation. Mo di cations expressing new system's needs range from minor c hanges to complete replacemen t of the con trol plan.
Page 7
name importance properties 1+ execCondition() execAction() 1+ activates updateBlackboard() updateControlPlan() Domain_KS Control_KS Knowledge Sources update() blackboardNodes access() Blackboard Control Plan Control reads reads planRepresentation access() update()

selectKS() configureKS() executeKS() operates on operates on 1+ Figure 4: System's structure The con trol comp onen t selects, con gures and executes kno wledge sources. Selection of the most appropri- ate kno wledge sources for execution is based on meta-con trol criteria con tained in the con trol plan. Structural relationships b et een these comp onen ts are summarized in Figure 4. Implemen tation The blac kb oard-based con trol mo del brings new comp onen ts that are not considered in the general approac h. In particular, dev elop ers ha e to deal with the implemen tation of con trol

plans. Con trol plans can b e implemen ted in man yw ys. The most general approac h is to implemen t a con trol plan as a graph. A no de represen ts a step in the problem solving pro cess and consists of an in tended activit y in the form of a tuple (task, parameters, constrain ts). Steps are link ed to eac h other through arcs lab eled with conditions ab out the problem solving state. The solving path is determined dynamically , considering the conditions lab eling the arcs of the graph. Con trol kno wledge sources are implemen ted as domain kno wledge sources in terms of in terfaces. In

terfaces comprise: condition of activ ation describing, among other p ossible things, the ev en ts that enable the kno wledge source, ariables to b e b ound in the enabling con text, Required resources, Non functional prop erties lik e safet y or eciency c haracterizing the execution, Results prop erties. The action part of con trol kno wledge source con tains either new plan that has to replace or precede the curren t one when the kno wledge source is executed, or a set of mo di cations to b e brough t to the curren t con trol plan. Kno wn Uses
Page 8
target Control Plan root

current Node constraint parameter Arc condition 1+ task source Figure 5: Con trol plan This pattern for con trol has b een used for ten ears in umerous domains, including autonomous rob ots [7], systems for monitoring medical patien ts [8 ], and pilot's asso ciate systems. Consequences This pattern presen ts the usual prop erties of blac kb oard systems as giv en in [1]. In particular, it pro vides framew ork in whic h appropriate sets of kno wledge sources can b e selected and con gured at b oth design time and run time. The in tegration of kno wledge sources is actually v ery simple since

kno wledge sources are considered indep enden tly and are only c haracterized b y their o wn prop erties. t run-time, if useful new application-relev an t kno wledge sources should b ecome a ailable, the new kno wledge sources can b e substituted for old ones or added to the kno wledge base alongside the old ones, without in terrupting system op erations. The arc hitecture's ev en t-based enabling of kno wledge sources, its plan-based meta-con trol hoices among comp eting kno wledge sources, and its e ort to retriev necessary kno wledge from the shared blac kb oard are not preprogrammed to

require an particular kno wledge sources. They op erate on whatev er kno wledge sources are a ailable in the kno wledge source library at run-time. In the other hand, the approac do es not guaran tee the deliv ering of a solution to a giv en problem, since appropriate kno wledge sources migh b e missing. Encapsulating meta-con trol elemen ts in single hangeable ob ject i.e. the con trol plan) brings considerable exibilit y in the system. Con trol strategies can b e adapted at run time to dynamic conditions. Th us, not only domain comp onen ts can b e plugged in the system b oth at run time and

design time, but con trol strategies as ell. Finally , this pattern is a v aluable basis for the design and dev elopmen t of application framew orks. Application framew orks are semi-complete applications pro viding arc hitectural foundations to solv e a range of domain-sp eci c problems and that need to b e sp ecialized to meet particular applications' requiremen ts. Suc h sp ecializations can b e done b y plugging the adequate domain kno wledge sources and con trol strategies in con trol kno wledge sources. See Also The PRS system ( Pr dur al R asoning System )[9 ] implemen ts a view of

dynamic con trol v ery close to the one presen ted in this pap er. PRS is a generic arc hitecture that can b oth react rapidly to unan ticipated c hanges in the en vironmen t and p erform goal-directed reasoning. o that end, PRS uses sp ecial plans, called Know le dge A as (KA), that include an in o cation condition whic h sp eci es whether the KA is useful and a b o dy whic h describ es the sequence of subgoals whic h constitutes the pro cedure. A subgoal can result in a primitiv e directly executable or in another KA in o cation. KA expresses sev eral ys to solv problem, dep ending on en

vironmen tal conditions and on the exact problem's nature. t eac h execution cycle, PRS c hec ks all the KAs for applicabilit and selects one for execution. The meta-con trol approac h of PRS is similar to the blac kb oard-based con trol and pro vides exibilit y and reac- tivit ormalism for kno wledge sources is restricted to the graph-based KA. Con trol plan corresp onds to the goal under consideration at eac h con trol cycle of PRS. PRS has b een applied in man y elds, including telecomm unication applications. Credits
Page 9
The blac kb oard-based con trol mo del has b een

designed and dev elop ed at Stanford Univ ersit yb y Barbara Ha es- Roth [10 ]. Ov er the y ears, sev eral exp erimen ts on real-w orld applications ha e b een conducted in order to v alidate the mo del and to sho w the wide v ariet y of domains that can b e tac kled with this approac h. Conclusion This pap er presen ts ery rst step to ards pattern language to de ne ulti-exp ert systems. Additional patterns w ould b e needed to describ e other existing implemen tations of the con trol comp onen t in a blac kb oard- based system lik e the hierarc hical con trol. atterns guiding the design of

kno wledge sources and the selection of appropriate tec hniques in a giv en situation w ould b e v ery useful as w ell. References [1] F. Busc hmann, R. Meunier, H. Rohnert, P . Sommerlad, M. Stal, Pattern-oriente d Softwar eA chite ctur e: System of Pattern , Wiley & Sons. [2] L. Erman, F. Ha es-Roth, V. Lesser, R. Reddy The He arsay-II sp ch understanding system: Inte gr ating know le dge to r esolve unc ertainty ,A CM Computing Surv eys 12 (2), 1980. [3] V. Lesser and D. Corkill, The Distribute ehicle Monitoring estb d: to ol for investigating distribute pr oblem solving networks , Arti

cial In telligence Magazine, v ol. 4, nos 3, 1983. [4] A. erry Using explicit str ate gic know le dge to ontr ol exp ert systems in Blac kb oard Systems, Addison- esley , 1988. [5] . Nii, Blackb ar d systems: Part I and II , The AI Magazine, v ol. 7, nos 2 and 3. [6] R. Engelmore, T. Morgan (Editors), Blackb ar d systems , Addison-W esley , 1988. [7] B. Ha es-Roth, P . Lalanda, P . Morignot, M. Balano vic and K. P eger, A domain-sp ci c softwar ear chite c- tur e for adaptative intel ligent system , IEEE T ransactions on Soft are Engineering, 1995. [8] B. Ha es-Roth and J.E. Larsson, domain-sp

ci c softwar ar chite ctur for class of intel ligent atient monitoring systems , Journal of Exp erimen tal and Theoretical Arti cial In telligence, 8(2), 1996. [9] M. George and F. Ingrand, De cision making in an emb dde asoning system In ternational Conference on Arti cial In telligence, IJCAI-89, 1989. [10] B. Ha es-Roth, A blackb ar dar chite ctur e for c ontr ol , Arti cial In telligence, n um. 26, 1985.