Decision Supp ort Systems Marek J PDF document - DocSlides

Decision Supp ort Systems Marek J PDF document - DocSlides

2014-12-12 144K 144 0 0


Druzdzel and Roger R Flynn Decision Systems Lab oratory Sc ho ol of Information Sciences and In telligen Systems Program Univ ersit of Pittsburgh Pittsburgh 15260 marekflynn sis p itt e du httpwwwsispi tt ed u ds app ear in Encyclop di ID: 22525

Direct Link: Embed code:

Download this pdf

DownloadNote - The PPT/PDF document "Decision Supp ort Systems Marek J" 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.

Presentations text content in Decision Supp ort Systems Marek J

Page 1
Decision Supp ort Systems Marek J. Druzdzel and Roger R. Flynn Decision Systems Lab oratory Sc ho ol of Information Sciences and In telligen Systems Program Univ ersit of Pittsburgh Pittsburgh, 15260 marek,flynn @sis .p itt .e du http://www.sis.pi tt .ed u/ ds app ear in Encyclop dia of Libr ary and Information Scienc Second Edition, Allen Ken (ed.), New ork: Marcel Dekk er, Inc., 2002
Page 2
Con ten ts In tro duction Decisions and Decision Mo deling yp es of Decisions Human Judgmen and Decision Making Mo deling Decisions Comp onen ts of Decision Mo dels Decision Supp ort Systems Normativ Systems Normativ and Descriptiv Approac hes Decision-Analytic Decision Supp ort Systems Equation-Based and Mixed Systems 10 User In terfaces to Decision Supp ort Systems 11 Supp ort for Mo del Construction and Mo del Analysis 11 Supp ort for Reasoning ab out the Problem Structure in Addition to Numerical Calculations 11 Supp ort for Both Choice and Optimization of Decision ariables 12 Graphical In terface 12 Summary 12
Page 3
In tro duction Making decisions concerning complex systems (e.g., the managemen of organizational op erations, industrial pro cesses, or in estmen ortfolios; the command and con trol of military units; or the con trol of uclear er plan ts) often strains our cognitiv capabilities. Ev en though individual in teractions among systemís ariables ma ell understo d, predicting ho the system will react to an external manipulation suc as olicy decision is often dicult. What will e, for example, the eect of in tro ducing the third shift on factory o or? One migh exp ect that this will increase the plan tís output roughly 50 ercen t. actors suc as additional ages, mac hine eardo wn, main tenance breaks, ra material usage, supply logistics, and future demand need also considered, ho ev er, as they all will impact the total nancial outcome of this decision. Man ariables are in olv ed in complex and often subtle in terdep endencies and predicting the total outcome ma daun ting. There is substan tial amoun of empirical evidence that uman in tuitiv judgmen and deci- sion making can far from optimal, and it deteriorates ev en further with complexit and stress. Because in man situations the qualit of decisions is imp ortan t, aiding the deciencies of uman judgmen and decision making has een ma jor fo cus of science throughout history Disciplines suc as statistics, economics, and op erations researc dev elop ed arious metho ds for making rational hoices. More recen tly these metho ds, often enhanced ariet of tec hniques originating from information science, cognitiv psyc hology and articial in telligence, ha een implemen ted in the form of computer programs, either as stand-alone to ols or as in tegrated computing en vironmen ts for complex decision making. Suc en vironmen ts are often giv en the common name of de cision supp ort systems (DSSs). The concept of DSS is extremely broad, and its denitions ary dep ending on the authorís oin of view. oid exclusion of an of the existing yp es of DSSs, will dene them roughly as in teractiv computer-based systems that aid users in judgmen and hoice activities. An- other name sometimes used as synon ym for DSS is know le dge-b ase systems whic refers to their attempt to formalize domain kno wledge so that it is amenable to mec hanized reasoning. Decision supp ort systems are gaining an increased opularit in arious domains, including busi- ness, engineering, the military and medicine. They are esp ecially aluable in situations in whic the amoun of ailable information is prohibitiv for the in tuition of an unaided uman decision mak er and in whic precision and optimalit are of imp ortance. Decision supp ort systems can aid uman cognitiv deciencies in tegrating arious sources of information, pro viding in telligen access to relev an kno wledge, and aiding the pro cess of structuring decisions. They can also supp ort hoice among ell-dened alternativ es and build on formal approac hes, suc as the metho ds of engineering economics, op erations researc h, statistics, and decision theory They can also emplo articial in tel- ligence metho ds to address heuristically problems that are in tractable formal tec hniques. Prop er application of decision-making to ols increases pro ductivit eciency and eectiv eness and giv es man businesses comparativ adv an tage er their comp etitors, allo wing them to mak optimal hoices for tec hnological pro cesses and their parameters, planning business op erations, logistics, or in estmen ts. While it is dicult to erestimate the imp ortance of arious computer-based to ols that are relev an to decision making (e.g., databases, planning soft are, and spreadsheets), this article fo cuses primarily on the core of DSS, the part that directly supp orts mo deling decision problems and iden ties est alternativ es. will briey discuss the haracteristics of decision problems and ho decision making can supp orted computer programs. then co er arious comp onen ts of DSSs and the role that they pla in decision supp ort. will also in tro duce an emergen class of normative systems (i.e., DSSs based on sound theoretical principles), and in particular, decision- analytic DSSs. Finally will review issues related to user in terfaces to DSSs and stress the imp ortance of user in terfaces to the ultimate qualit of decisions aided computer programs.
Page 4
Decisions and Decision Mo deling yp es of Decisions simple view of decision making is that it is problem of hoice among sev eral alternativ es. somewhat more sophisticated view includes the pro cess of constructing the alternativ es (i.e., giv en problem statemen t, dev eloping list of hoice options). complete picture includes searc for opp ortunities for decisions (i.e., disco ering that there is decision to made). manager of compan ma face hoice in whic the options are clear (e.g., the hoice of supplier from among all existing suppliers). She ma also face ell-dened problem for whic she designs creativ decision options (e.g., ho to mark et new pro duct so that the prots are maximized). Finally she ma ork in less reactiv fashion and view decision problems as opp ortunities that ha to disco ered studying the op erations of her compan and its surrounding en vironmen (e.g., ho can she mak the pro duction pro cess more ecien t). There is uc anecdotal and some empirical evidence that structuring decision problems and iden tifying creativ decision alternativ es determine the ultimate qualit of decisions. Decision supp ort systems aim mainly at this broadest yp of decision making, and in addition to supp orting hoice, they aid in mo deling and analyzing systems (suc as complex organizations), iden tifying decision opp ortunities, and structuring decision problems. Human Judgmen and Decision Making Theoretical studies on rational decision making, notably that in the con text of probabilit theory and decision theory ha een accompanied empirical researc on whether uman eha vior complies with the theory It has een rather con vincingly demonstrated in umerous empirical studies that uman judgmen and decision making is based on in tuitiv strategies as opp osed to theoretically sound reasoning rules. These in tuitiv strategies, referred to as judgmental heuristics in the con text of decision making, help us in reducing the cognitiv load, but alas at the exp ense of optimal decision making. Eectiv ely our unaided judgmen and hoice exhibit systematic violations of probabilit axioms (referred to as biases ). ormal discussion of the most imp ortan researc results along with exp erimen tal data can found in an an thology edited Kahneman, Slo vic, and Tv ersky [16 ]. Da es [2 pro vides an accessible in tro duction to what is kno wn ab out eopleís decision-making erformance. One migh hop that eople who ha ac hiev ed exp ertise in domain will not sub ject to judgmen tal biases and will approac optimalit in decision making. While empirical evidence sho ws that exp erts indeed are more accurate than no vices within their area of exp ertise, it also sho ws that they also are liable to the same judgmen tal biases as no vices and demonstrate apparen errors and inconsistencies in their judgmen t. Professionals suc as practicing ph ysicians use essen tially the same judgmen tal heuristics and are prone to the same biases, although the degree of departure from the normativ ely prescrib ed judgmen seems to decrease with exp erience. In addition to lab oratory evidence, there are sev eral studies of exp ert erformance in realistic settings, sho wing that it is inferior ev en to simple linear mo dels (an informal review of the ailable evidence and oin ters to literature can found in the ok Da es [2]). or example, predictions of future violen eha vior of psyc hiatric patien ts made panel of psyc hiatrists who had access to patien records and in terview ed the patien ts ere found to inferior to simple mo del that included only the past incidence of violen eha vior. Predictions of marriage counselors concerning marital happiness ere sho wn to inferior to simple mo del that just subtracted the rate of gh ting from the rate of sexual in tercourse (again, the marriage counselors had access to all data, including in terviews with the couples). Studies yielding similar results ha een conducted with bank loan ocers, ph ysicians, univ ersit admission committees, and so on.
Page 5
Mo deling Decisions The sup eriorit of ev en simple linear mo dels er uman in tuitiv judgmen suggests that one to impro the qualit of decisions is to decomp ose decision problem in to simpler comp onen ts that are ell dened and ell understo d. Studying complex system built out of suc comp onen ts can subsequen tly aided formal, theoretically sound tec hnique. The pro cess of decomp osing and formalizing problem is often called mo deling. Mo deling amoun ts to nding an abstract represen tation of real-w orld system that simplies and assumes as uc as ossible ab out the system, and while retaining the systemís essen tial relationships, omits unnecessary detail. Building mo del of decision problem, as opp osed to reasoning ab out problem in holistic allo ws for applying scien tic kno wledge that can transferred across problems and often across domains. It allo ws for analyzing, explaining, and arguing ab out decision problem. The desire to impro uman decision making pro vided motiv ation for the dev elopmen of ariet of mo deling to ols in disciplines of economics, op erations researc h, decision theory decision analysis, and statistics. In eac of these mo deling to ols, kno wledge ab out system is represen ted means of algebraic, logical, or statistical ariables. In teractions among these ariables are expressed equations or logical rules, ossibly enhanced with an explicit represen tation of uncertain When the functional form of an in teraction is unkno wn, it is sometimes describ ed in purely probabilistic terms; for example, conditional probabilit distribution. Once mo del has een form ulated, ariet of mathematical metho ds can used to analyze it. Decision making under certain has een addressed economic and op erations researc metho ds, suc as cash o analysis, break- ev en analysis, scenario analysis, mathematical programming, in en tory tec hniques, and ariet of optimization algorithms for sc heduling and logistics. Decision making under uncertain enhances the ab metho ds with statistical approac hes, suc as reliabilit analysis, sim ulation, and statistical decision making. Most of these metho ds ha made it in to college curricula and can found in managemen textb oks. Due to space constrain ts, will not discuss their details further. Comp onen ts of Decision Mo dels While mathematically mo del consists of ariables and sp ecication of in teractions among them, from the oin of view of decision making mo del and its ariables represen the follo wing three comp onen ts: measure of preferences er decision ob jectiv es, ailable decision options, and measure of uncertain er ariables inuencing the decision and the outcomes. Preference is widely view ed as the most imp ortan concept in decision making. Outcomes of decision pro cess are not all equally attractiv and it is crucial for decision mak er to examine these outcomes in terms of their desirabilit Preferences can ordinal (e.g., more income is preferred to less income), but it is con enien and often necessary to represen them as umerical quan tities, esp ecially if the outcome of the decision pro cess consists of ultiple attributes that need to compared on common scale. Ev en when they consist of just single attribute but the hoice is made under uncertain expressing preferences umerically allo ws for trade-os et een desirabilit and risk. The second comp onen of decision problems is ailable decision options. Often these options can en umerated (e.g., list of ossible suppliers), but sometimes they are con tin uous alues of sp ecied olicy ariables (e.g., the amoun of ra material to ept in sto k). Listing the ailable decision options is an imp ortan elemen of mo del structuring. The third elemen of decision mo dels is uncertain Uncertain is one of the most inheren and most prev alen prop erties of kno wledge, originating from incompleteness of information, imprecision,
Page 6
and mo del appro ximations made for the sak of simplicit It ould not an exaggeration to state that real-w orld decisions not in olving uncertain either do not exist or elong to truly limited class. Decision making under uncertain can view ed as delib eration: determining what action should tak en that will maximize the exp ected gain. Due to uncertain there is no guaran tee that the result of the action will the one in tended, and the est one can hop for is to maximize the hance of desirable outcome. The pro cess rests on the assumption that go decision is one that results from go decision-making pro cess that considers all imp ortan factors and is explicit ab out decision alternativ es, preferences, and uncertain It is imp ortan to distinguish et een go decisions and go outcomes. By strok of go luc or decision can lead to ery go outcome. Similarly ery go decision can follo ed bad outcome. Supp orting decisions means supp orting the decision-making pro cess so that etter decisions are made. Better decisions can exp ected to lead to etter outcomes. Decision Supp ort Systems Decision supp ort systems are in teractiv e, computer-based systems that aid users in judgmen and hoice activities. They pro vide data storage and retriev al but enhance the traditional information access and retriev al functions with supp ort for mo del building and mo del-based reasoning. They supp ort framing, mo deling, and problem solving. ypical application areas of DSSs are managemen and planning in business, health care, the military and an area in whic managemen will encoun ter complex decision situations. Deci- sion supp ort systems are ypically used for strategic and tactical decisions faced upp er-lev el managemen t|decisions with reasonably lo frequency and high oten tial consequences|in whic the time tak en for thinking through and mo deling the problem pa ys o generously in the long run. There are three fundamen tal comp onen ts of DSSs [22 ]. Datab ase management system (DBMS). DBMS serv es as data bank for the DSS. It stores large quan tities of data that are relev an to the class of problems for whic the DSS has een designed and pro vides logical data structures (as opp osed to the ph ysical data structures) with whic the users in teract. DBMS separates the users from the ph ysical asp ects of the database structure and pro cessing. It should also capable of informing the user of the yp es of data that are ailable and ho to gain access to them. Mo del-b ase management system (MBMS). The role of MBMS is analogous to that of DBMS. Its primary function is pro viding indep endence et een sp ecic mo dels that are used in DSS from the applications that use them. The purp ose of an MBMS is to transform data from the DBMS in to information that is useful in decision making. Since man problems that the user of DSS will cop with ma unstructured, the MBMS should also capable of assisting the user in mo del building. Dialo gener ation and management system (DGMS). The main pro duct of an in teraction with DSS is insigh t. As their users are often managers who are not computer-trained, DSSs need to equipp ed with in tuitiv and easy-to-use in terfaces. These in terfaces aid in mo del As Benjamin ranklin expressed it in 1789 in letter to his friend M. Le Ro \in this orld nothing can said to certain, except death and taxes" The Complete Works of Benjamin anklin John Bigelo (ed), New ork and London: G.P Putnamís Sons, 1887, ol. 10, page 170).
Page 7
building, but also in in teraction with the mo del, suc as gaining insigh and recommendations from it. The primary resp onsibilit of DGMS is to enhance the abilit of the system user to utilize and enet from the DSS. In the remainder of this article, will use the broader term user interfac rather than DGMS. While ariet of DSSs exists, the ab three comp onen ts can found in man DSS arc hitectures and pla prominen role in their structure. In teraction among them is illustrated in Fig. 1. Essen tially the user in teracts with the DSS through the DGMS. This comm unicates with the DBMS Model Base abase MBMS DBMS DGMS DSS User Figure 1: The arc hitecture of DSSs (after Sage, Ref. [22 ]). and MBMS, whic screen the user and the user in terface from the ph ysical details of the mo del base and database implemen tation. Normativ Systems Normativ and Descriptiv Approac hes Whether or not one trusts the qualit of uman in tuitiv reasoning strategies has profound im- pact on oneís view of the philosophical and tec hnical foundations of DSSs. There are distinct approac hes to supp orting decision making. The rst aims at building supp ort pro cedures or systems that imitate uman exp erts. The most prominen mem er of this class of DSSs are exp ert systems computer programs based on rules elicited from uman domain exp erts that imitate reasoning of uman exp ert in giv en domain. Exp ert systems are often capable of supp orting decision making in that domain at lev el comparable to uman exp erts. While they are exible and often able to address complex decision problems, they are based on in tuitiv uman reasoning and lac soundness and formal guaran tees with resp ect to the theoretical reliabilit of their results. The danger of the exp ert system approac h, increasingly appreciated DSS builders, is that along with imitating uman thinking and its ecien heuristic principles, ma also imitate its undesirable a ws [13 ]. The second approac is based on the assumption that the most reliable metho of dealing with complex decisions is through small set of normativ ely sound principles of ho decisions should made. While heuristic metho ds and ad ho reasoning sc hemes that imitate uman cognition ma in man domains erform ell, most decision mak ers will reluctan to rely on them whenev er the cost of making an error is high. giv an extreme example, few eople ould ho ose to y airplanes built using heuristic principles er airplanes built using the la ws of aero dynamics enhanced with probabilistic reliabilit analysis. Application of formal metho ds in DSSs mak es these systems
Page 8
philosophically distinct from those based on ad ho heuristic articial in telligence metho ds, suc as rule-based systems. The goal of DSS, according to this view, is to supp ort unaided uman in tuition, just as the goal of using calculator is to aid umanís limited capacit for men tal arithmetic. Decision-Analytic Decision Supp ort Systems An emergen class of DSSs kno wn as de cision-analytic DSSs applies the principles of decision theory probabilit theory and decision analysis to their decision mo dels. Decision theory is an axiomatic theory of decision making that is built on small set of axioms of rational decision making. It expresses uncertain in terms of probabilities and preferences in terms of utilities. These are com- bined using the op eration of mathematical exp ectation. The attractiv eness of probabilit theory as formalism for handling uncertain in DSSs, lies in its soundness and its guaran tees concerning long-term erformance. Probabilit theory is often view ed as the gold standard for rationalit in reasoning under uncertain ollo wing its axioms oers protection from some elemen tary inconsis- tencies. Their violation, on the other hand, can demonstrated to lead to sure losses [23 ]. Decision analysis is the art and science of applying decision theory to real-w orld problems. It includes ealth of tec hniques for mo del construction, suc as metho ds for elicitation of mo del structure and probabilit distributions that allo minimization of uman bias, metho ds for hec king the sensitivit of mo del to imprecision in the data, computing the alue of obtaining additional information, and presen tation of results. (See, for example, Ref. [27 for basic review of the ailable tec hniques.) These metho ds ha een under con tin uous scrutin psyc hologists orking in the domain of e- ha vioral decision theory and ha pro en to cop reasonably ell with the dangers related to uman judgmen tal biases. Normativ systems are usually based on graphical probabilistic mo dels, whic are represen tations of the join probabilit distribution er mo delís ariables in terms of directed graphs. Directed graphs, suc as the one in Fig. 2, are kno wn as Ba esian net orks (BNs) or causal net orks [19 ]. Ba esian net orks oer compact represen tation of join probabilit distributions and are capable of practical represen tation of large mo dels, consisting of tens or undreds of ariables. Ba esian net orks can easily extended with decision and alue ariables for mo deling decision problems. The former denote ariables that are under the decision mak erís con trol and can directly ma- nipulated, and the latter enco de users preferences er arious outcomes of the decision pro cess. Suc amended graphs are kno wn as inuenc diagr ams [15 ]. Both the structure and the umerical probabilit distributions in BN can elicited from uman exp ert and are reection of the exp ertís sub jectiv view of real-w orld system. If ailable, scien tic kno wledge ab out the system, oth in terms of the structure and frequency data, can easily incorp orated in the mo del. Once mo del has een created, it is optimized using formal decision-theoretic algorithms. Decision anal- ysis is based on the empirically tested paradigm that eople are able to reliably store and retriev their ersonal eliefs ab out uncertain and preferences for dieren outcomes, but are uc less reliable in aggregating these fragmen ts in to global inference. While uman exp erts are excellen in structuring problem, determining the comp onen ts that are relev an to it and pro viding lo cal estimates of probabilities and preferences, they are not reliable in com bining man simple factors in to an optimal decision. The role of decision-analytic DSS is to supp ort them in their eaknesses using the formal and theoretically sound principles of statistics. The approac tak en decision analysis is compatible with that of DSSs. The goal of decision analysis is to pro vide insigh in to decision. This insigh t, consisting of the analysis of all relev an factors, their uncertain and the critical nature of some assumptions, is ev en more imp ortan than the actual recommendation. Decision-analytic DSSs ha een successfully applied to practical systems in medicine, business,
Page 9
Figure 2: Example of Ba esian net ork mo deling teac hing exp enditures in univ ersit op erations. and engineering. As these systems tend to naturally ev olv in to three not necessarily distinct classes, it ma in teresting to compare their structure and arc hitectural organization. Systems with static domain mo dels. In this class of systems, probabilistic domain is rep- resen ted large net ork enco ding the domainís structure and its umerical parameters. The net ork comprising the domain mo del is normally built decision analysts and domain exp erts. An example migh medical diagnostic system co ering certain class of disor- ders. Queries in suc system are answ ered assigning alues to those no des of the net ork that constitute the observ ations for particular case and propagating the impact of the ob- serv ation through the net ork in order to nd the probabilit distribution of some selected no des of in terest (e.g., no des that represen diseases). Suc net ork can, on case-b y-case basis, extended with decision no des and alue no des to supp ort decisions. Systems with static domain mo dels are conceptually similar to rule-based exp ert systems co ering an area of exp ertise. Systems with customize de cision mo dels. The main idea ehind this approac is automatic generation of graphical decision mo del on er-case basis in an in teractiv eort et een the DSS and the decision mak er. The DSS has domain exp ertise in certain area and pla ys the role of decision analyst. During this in teraction, the program creates customized inuence diagram, whic is later used for generating advice. The main motiv ation for this approac is the premise that ev ery decision is unique and needs to lo ok ed at individually; an inuence diagram needs to tailored to individual needs [14 ]. Some examples of applications are describ ed in sp ecial issue of Communic ations of the CM on practical applications of decision-theoretic metho ds (v ol. 38, no. 3, Marc 1995). The readers can exp erimen with GeNIe [7], dev elopmen system for decision-analytic DSSs dev elop ed at the Decision Systems Lab oratory Univ ersit of Pittsburgh, ailable at http://www2.sis.pitt.ed u/ enie
Page 10
Systems ap able of le arning mo del fr om data The third class of systems emplo ys computer- in tensiv statistical metho ds for learning mo dels from data [1, 11 12 21 26 ]. Whenev er there are sucien data ailable, the systems can literally learn graphical mo del from these data. This mo del can subsequen tly used to supp ort decisions within the same domain. The rst approac hes are suited for sligh tly dieren applications. The customized mo del gener- ation approac is an attempt to automate the most lab orious part of decision making, structuring problem, so far done with signican assistance from trained decision analysts. session with the program that assists the decision mak er in building an inuence diagram is lab orious. This mak es the customized mo del generation approac particularly suitable for decision problems that are infre- quen and serious enough to treated individually Because in the static domain mo del approac an existing domain mo del needs to customized the case data only the decision-making cycle is rather short. This mak es it particularly suitable for those decisions that are highly rep etitiv and need to made under time constrain ts. practical system can com bine the three approac hes. static domain mo del can sligh tly customized for case that needs individual treatmen t. Once completed, customized mo del can blended in to the large static mo del. Learning systems can supp ort oth the static and the customized mo del approac h. On the other hand, the learning pro cess can greatly enhanced prior kno wledge from domain exp erts or prior mo del. Equation-Based and Mixed Systems In man business and engineering problems, in teractions among mo del ariables can describ ed equations whic h, when solv ed sim ultaneously can used to predict the eect of decisions on the system, and hence supp ort decision making. One sp ecial yp of sim ultaneous equation mo del is kno wn as the structural equation mo del (SEM), whic has een opular metho of represen ting systems in econometrics. An equation is structural if it describ es unique, indep enden causal mec hanism acting in the system. Structural equations are based on exp ert kno wledge of the system com bined with theoretical considerations. Structural equations allo for natural, mo dular description of system|eac equation represen ts its individual comp onen t, separable and indep enden mec hanism acting in the system|y et, the main adv an tage of ha ving structural mo del is, as explicated Simon [24 ], that it includes causal information and aids predictions of the eects of external in terv en tions. In addition, the causal structure of structural equation mo del can represen ted graphically [24], whic allo ws for com bining them with decision-analytic graphical mo dels in practical systems [9, 20 ]. Structural equation mo dels oer signican adv an tages for olicy making. Often decision mak er confron ted with complex system needs to decide not only the alues of olicy ariables but also whic ariables should manipulated. hange in the set of olicy ariables has profound impact on the structure of the problem and on ho their alues will propagate through the system. The user determines whic ariables are olicy ariables and whic are determined within the mo del. hange in the SEMs or the set of olicy ariables can reected rapid restructuring of the mo del and predictions in olving this new structure [25 ]. Our long-term pro ject, the En vironmen for Strategic Planning (ESP) [6], is based on ybrid graphical mo deling to ol that com bines SEMs with decision-analytic principles. ESP is capable of represen ting oth discrete and con tin uous ariables in olv ed in deterministic and probabilistic relationships. The erful features of SEMs allo ESP to act as graphical spreadsheet in tegrating umerical and sym olic metho ds and allo wing the indep enden ariables to selected at will without ha ving to reform ulate the mo del eac time. This pro vides an immense exibilit that is not aorded 10
Page 11
ordinary spreadsheets in ev aluating alternate olicy options. User In terfaces to Decision Supp ort Systems While the qualit and reliabilit of mo deling to ols and the in ternal arc hitectures of DSSs are imp or- tan t, the most crucial asp ect of DSSs is, far, their user in terface. Systems with user in terfaces that are cum ersome or unclear or that require un usual skills are rarely useful and accepted in practice. The most imp ortan result of session with DSS is insigh in to the decision problem. In addition, when the system is based on normativ principles, it can pla tutoring role; one migh hop that users will learn the domain mo del and ho to reason with it er time, and impro their wn thinking. go user in terface to DSSs should supp ort mo del construction and mo del analysis, reasoning ab out the problem structure in addition to umerical calculations and oth hoice and optimization of decision ariables. will discuss these in the follo wing sections. Supp ort for Mo del Construction and Mo del Analysis User in terface is the ehicle for oth mo del construction (or mo del hoice) and for in estigating the results. Ev en if system is based on theoretically sound reasoning sc heme, its recommendations will as go as the mo del they are based on. urthermore, ev en if the mo del is ery go appro ximation of realit and its recommendations are correct, they will not follo ed if they are not understo d. Without understanding, the users ma accept or reject systemís advice for the wrong reasons and the com bined decision-making erformance ma deteriorate ev en elo unaided erformance [17 ]. go user in terface should mak the mo del on whic the systemís reasoning is based transparen to the user. Mo deling is rarely one-shot pro cess, and go mo dels are usually rened and enhanced as their users gather practical exp eriences with the system recommendations. It is imp ortan to strik careful balance et een precision and mo deling eorts; some parts of mo del need to ery precise while others do not. go user in terface should include to ols for examining the mo del and iden tifying its most sensitiv parts, whic can subsequen tly elab orated on. Systems emplo ed in practice will need their mo dels rened, and go user in terface should mak it easy to access, examine, and rene its mo dels. Some oin ters to ork on supp ort for building decision-analytic systems can found in [8, 10 18 28 ]. Supp ort for Reasoning ab out the Problem Structure in Addition to Nu- merical Calculations While umerical calculations are imp ortan in decision supp ort, reasoning ab out the problem struc- ture is ev en more imp ortan t. Often when the system and its mo del are complex it is insigh tful for the decision mak er to realize ho the system ariables are in terrelated. This is helpful in designing creativ decision options but also in understanding ho olicy decision will impact the ob jectiv e. Graphical mo dels, suc as those used in decision analysis or in equation-based and ybrid sys- tems, are particularly suitable for reasoning ab out structure. Under certain assumptions, directed graphical mo del can giv en causal in terpretation. This is esp ecially con enien in situations where the DSS autonomically suggests decision options; giv en causal in terpretation of its mo del, 11
Page 12
it is capable of predicting eects of in terv en tions. causal graph facilitates building an eectiv user in terface. The system can refer to causal in teractions during its dialogue with the user, whic is kno wn to enhance user insigh [3]. Supp ort for Both Choice and Optimization of Decision ariables Man DSSs ha an inexible structure in the sense that the ariables that will manipulated are determined at the mo del-building stage. This is not ery suitable for planning of the strategic yp when the ob ject of the decision-making pro cess is iden tifying oth the ob jectiv es and the metho ds of ac hieving them. or example, hanging olicy ariables in spreadsheet-based mo del often requires that the en tire spreadsheet rebuilt. If there is no supp ort for that, few users will consider it as an option. This closes the orld of ossibilities for exible reframing of decision problem in the exploratory pro cess of searc hing for opp ortunities. Supp ort for oth hoice and optimization of decision ariables should an inheren part of DSSs. Graphical In terface Insigh in to mo del can increased greatly at the user in terface lev el diagram represen ting the in teractions among its comp onen ts; for example, dra wing of graph on whic mo del is based, suc as in Fig. 2. This graph is qualitativ e, structural explanation of ho information o ws from the indep enden ariables to the dep enden ariables of in terest. As mo dels ma ecome ery large, it is con enien to structure them in to submo dels, groups of ariables that form subsystem of the mo deled system. Suc submo dels can again sho wn graphically with in teractions among them, increasing simplicit and clarit of the in terface. Fig. sho ws submo del-lev el view of mo del dev elop ed in our ESP pro ject. Note that the graph in Fig. is an expanded ersion of the aching Exp enditur es submo del in Fig. 3. The user can na vigate through the hierarc of the en tire mo del in her quest for insigh t, op ening and closing submo dels on demand. Some oin ters to ork on user in terfaces of decision-analytic systems can found in [4, 28 ]. Summary Decision supp ort systems are erful to ols in tegrating scien tic metho ds for supp orting complex decisions with tec hniques dev elop ed in information science, and are gaining an increased opularit in man domains. They are esp ecially aluable in situations in whic the amoun of ailable infor- mation is prohibitiv for the in tuition of an unaided uman decision mak er and in whic precision and optimalit are of imp ortance. Decision supp ort systems aid uman cognitiv deciencies in tegrating arious sources of information, pro viding in telligen access to relev an kno wledge, aiding the pro cess of structuring, and optimizing decisions. Normativ DSSs oer theoretically correct and app ealing of handling uncertain and preferences in decision problems. They are based on carefully studied empirical principles underlying the discipline of decision analysis and they ha een successfully applied in man practical systems. eliev that they oer sev eral attractiv features that are lik ely to prev ail in the long run as far as the tec hnical dev elopmen ts are concerned. Because DSSs do not replace umans but rather augmen their limited capacit to deal with complex problems, their user in terfaces are critical. The user in terface determines whether DSS 12
Page 13
Figure 3: submo del-lev el view of decision mo del. will used at all and if so, whether the ultimate qualit of decisions will higher than that of an unaided decision mak er. Ac kno wledgmen ts ork on this article as supp orted the National Science oundation under acult Early Career Dev elopmen (CAREER) Program, gran IRI{9624629, the Air orce Oce of Scien tic Researc under gran ts F49620{97{1{0225 and F49620{00{1{011 2, and the Univ ersit of Pittsburgh Cen tral Researc Dev elopmen und. Figures and are snapshots of GeNIe, general purp ose dev elopmen en vironmen for graphical decision supp ort systems dev elop ed the Decision Systems Lab oratory Univ ersit of Pittsburgh and ailable at http://www.sis.p it t.e du en ie ould lik to thank Ms. Nanette urcik for her assistance with tec hnical editing. References [1] Gregory F. Co op er and Edw ard Hersk vits. Ba esian metho for the induction of probabilistic net orks from data. Machine arning 9(4):309{347, 1992. [2] Rob yn M. Da es. ational Choic in an Unc ertain World Hartcourt Brace Jo ano vic h, Pub- lishers, 1988. [3] Marek J. Druzdzel. Pr ob abilistic asoning in De cision Supp ort Systems: om Computation to Common Sense PhD thesis, Departmen of Engineering and Public olicy Carnegie Mellon Univ ersit Pittsburgh, A, Decem er 1992. 13
Page 14
[4] Marek J. Druzdzel. Explanation in probabilistic systems: Is it feasible? will it ork? In Pr dings of the Fifth International Workshop on Intel ligent Information Systems (WIS{96) pages 12{24, D eblin, oland, 2{5 JUne 1996. [5] Marek J. Druzdzel. Fiv useful prop erties of probabilistic kno wledge represen tations from the oin of view of in telligen systems. undamenta Informatic, Sp cial issue on Know le dge epr esentation and Machine arning 30(3/4):241{254, June 1997. [6] Marek J. Druzdzel. ESP: mixed initiativ decision-theoretic decision mo deling system. In Working Notes of the AAAI{99 Workshop on Mixe d-initiative Intel ligenc pages 99{106, Or- lando, FL, 18 July 1999. [7] Marek J. Druzdzel. SMILE: Structural Mo deling, Inference, and Learning Engine and GeNIe: dev elopmen en vironmen for graphical decision-theoretic mo dels. In Pr dings of the Six- te enth National Confer enc on rticial Intel ligenc (AAAI{99) pages 902{903, Orlando, FL, July 18{22 1999. [8] Marek J. Druzdzel and F. Ja vier ez. Criteria for com bining kno wledge from dieren sources in probabilistic mo dels. In Working Notes of the workshop on usion of Domain Kno wledge with Data for Decision Supp ort Sixte enth nnual Confer enc on Unc ertainty in rticial In- tel ligenc (UAI{2000) pages 23{29, Stanford, CA, 30 June 2000. [9] Marek J. Druzdzel and Herb ert A. Simon. Causalit in Ba esian elief net orks. In Pr dings of the Ninth nnual Confer enc on Unc ertainty in rticial Intel ligenc (UAI{93) pages 3{11, San rancisco, CA, 1993. Morgan Kaufmann Publishers, Inc. [10] Marek J. Druzdzel and Linda C. an der Gaag. Building probabilistic net orks: \Where do the um ers come from?" guest editors in tro duction. IEEE ansactions on Know le dge and Data Engine ering 12(4):481{486, July{August 2000. [11] Clark Glymour and Gregory F. Co op er, editors. Computation, Causation, and Disc overy AAAI Press, Menlo ark, CA, 1999. [12] Da vid E. Hec erman, Dan Geiger, and Da vid M. Chic ering. Learning Ba esian net orks: The com bination of kno wledge and statistical data. Machine arning 20(3):197{243, 1995. [13] Max Henrion, John S. Breese, and Eric J. Horvitz. Decision Analysis and Exp ert Systems. AI Magazine 12(4):64{91, Win ter 1991. [14] Sam uel Holtzman. Intel ligent De cision Systems Addison-W esley Reading, MA, 1989. [15] Ronald A. Ho ard and James E. Matheson. Inuence diagrams. In Ronald A. Ho ard and James E. Matheson, editors, The Principles and Applic ations of De cision nalysis pages 719{ 762. Strategic Decisions Group, Menlo ark, CA, 1984. [16] Daniel Kahneman, aul Slo vic, and Amos Tv ersky editors. Judgment Under Unc ertainty: Heuristics and Biases Cam bridge Univ ersit Press, Cam bridge, 1982. [17] aul E. Lehner, Theresa M. Mullin, and Marvin S. Cohen. probabilit analysis of the usefulness of decision aids. In M. Henrion, R.D. Shac ter, L.N. Kanal, and J.F. Lemmer, editors, Unc ertainty in rticial Intel ligenc pages 427{436. Elsevier Science Publishers B.V. (North Holland), 1990. [18] Tsai-Ching Lu, Marek J. Druzdzel, and Tze-Y un Leong. Causal mec hanism-based mo del con- struction. In Pr dings of the Sixte enth nnual Confer enc on Unc ertainty in rticial Intel- ligenc (UAI{2000) pages 353{362, San rancisco, CA, 2000. Morgan Kaufmann Publishers, Inc. 14
Page 15
[19] Judea earl. Pr ob abilistic asoning in Intel ligent Systems: Networks of Plausible Infer enc Morgan Kaufmann Publishers, Inc., San Mateo, CA, 1988. [20] Judea earl. Causality: Mo dels, asoning, and Infer enc Cam bridge Univ ersit Press, Cam- bridge, UK, 2000. [21] Judea earl and Thomas S. erma. theory of inferred causation. In J.A. Allen, R. Fik es, and E. Sandew all, editors, KR{91, Principles of Know le dge epr esentation and asoning: Pr dings of the Se ond International Confer enc pages 441{452, Cam bridge, MA, 1991. Morgan Kaufmann Publishers, Inc., San Mateo, CA. [22] Andrew Sage. De cision Supp ort Systems Engine ering John Wiley Sons, Inc., New ork, 1991. [23] Leonard J. Sa age. The oundations of Statistics (Se ond evise Edition) Do er Publications, New ork, NY, 1972. [24] Herb ert A. Simon. Causal ordering and iden tiabilit In William C. Ho and Tjalling C. Ko opmans, editors, Studies in Ec onometric Metho d. Cow les Commission for ese ar ch in Ec o- nomics. Mono gr aph No. 14 hapter I, pages 49{74. John Wiley Sons, Inc., New ork, NY, 1953. [25] Herb ert A. Simon, Ja an R. Kalagnanam, and Marek J. Druzdzel. erformance budget plan- ning: The case of researc univ ersit In preparation, 2000. [26] eter Spirtes, Clark Glymour, and Ric hard Sc heines. Causation, Pr diction, and Se ar ch Springer erlag, New ork, 1993. [27] Detlof on Win terfeldt and ard Edw ards. De cision nalysis and Behavior al ese ar ch Cam- bridge Univ ersit Press, Cam bridge, 1988. [28] Haiqin ang and Marek J. Druzdzel. User in terface to ols for na vigation in conditional proba- bilit tables and elicitation of probabilities in Ba esian net orks. In Pr dings of the Sixte enth nnual Confer enc on Unc ertainty in rticial Intel ligenc (UAI{2000) pages 617{625, San rancisco, CA, 2000. Morgan Kaufmann Publishers, Inc. 15

About DocSlides
DocSlides allows users to easily upload and share presentations, PDF documents, and images.Share your documents with the world , watch,share and upload any time you want. How can you benefit from using DocSlides? DocSlides consists documents from individuals and organizations on topics ranging from technology and business to travel, health, and education. Find and search for what interests you, and learn from people and more. You can also download DocSlides to read or reference later.