{EEE TRANSACTIONS ON SYSTEMS, MAN, ANO CYBERNETICS, VOL, sMC-10, NO. 12, DECEMBER 1980 I 1 Structural Modeling—A Tutorial Guide GEORGE G. LENDARIS, SENIOR MEMBER, TEEE Abstroct—Structural modeling (SM) refers to thave modellag activites tn whlch the Intention of the modeler Isto embody the geomet eather than the algebrale, and of describing form ruber than calculating oF measuring quantitadve output. A structural model a diagram which ‘onal ofa set of nodes tod connections between the oodes. A number of tools hare bec developed to asist In building and analyzing structural ‘models, An overview of selected structural modeling tools Is presented, ‘The overview is summarized as 2 diagram which is itself x strocunl ‘wodel This “map” portrays in an orgaalzed way the relationships between 1) the basis of digraph theory (hich underiies all the SM tool), 2) the extended assumpcons and data required by the selected SM tools, and 3) the types of Information that can be obtained sbout the systems Delng modeled In this way. Its a coupler map, but x step-by-step easy-to-follow procedure ts wed for slg the reader throug Its development. The paper i intended foc both fhe poteatial and exlsing users of the SM metindology. A number of tmpzestioas are gives: some praccal oves for the new wer and some ‘esearcvorlented ones for those Interested in advancing the SM capabil- ‘Hes applicability. I. Intropucrion TRUCTURAL MODELING (SM) is a process gen- erally involving participation by more than one per- son. It applies in cases where the participants are working collectively on a problem, and the problem is defined in terms of a system (elements, interconnections, etc.) being considered. The process starts with certain systemvrelated data, ideas, skills; and for knowledge residing in the vari- ous participants, and ends with an enhanced understand ing of the system by the participants, individually and collectively. ‘A number of methodologies have been developed to guide the structural modeling process. A survey of these methodologies (herein called SM tools) was recently car- ried out as part of a larger project [63]. The purpose of this paper is to share the insights gained during that study and the results of the comparative evaluations that were made, ‘The material of this paper should be of interest to the initiated user, as well as to the potential user of the SM methodologies: to the potential user because it is an organized introduction to an emerging field and to the initiated user because it provides 1) an introduction to some SM tools other than those likely being used, and 2) an overview of how the selected tools relate to one ‘Manuscript received Apeil 1977; revised: August 15, 1980. This paper was adapied from the Final Report of & project sponsored by the ‘Natioedl Science Foundation under Grant ERS 7621041. "The author is wih the Systems Science Ph.D. Program, Portland State University, P.O. Box 751, Pordand, OR 97207. another, and how they relate to the mathematical theory of digraphs, upon which they are all based, The SM tools studied have as a common basis the mathematical theory for directed graphs as well as the related structural modeling concepts developed by Harary et al, (38], Roberts [81], and Warfield (102). The basic mathematical entity common to these tools is called a structural model. A structural mode! is simply a collection of elements and their relationships. Graphically these models are represented by a set of nodes with some or all the nodes connected by lines. 'A key aspect of structural modeling versus modeling without the adjective structural is one of emphasis; em- phasis on qualitative structural (geometric, topological, etc.) aspects rather than on exact aumerical or statistical properties of the systems being modeled. However itis not sufficient to say just that structural models are geometric and qualitative; arithmetic operations and quantitative output are encountered in some implementations. But “geometric” and “qualitative” are important adjectives, indicating as they do that structural models are in a class apart from those models more widely familiar in techno- logical fields. Misunderstanding of these distinctions could lead to potential misuse of structural models for predictive purposes. For example, an uncalibrated dynamic model (Gee Section II, Output [IIB) whose purpose is to provide a qualitative sense of the system's behavior could be (mis)used to make specific time predictions. Structural modeling holds the promise of converting 2 completely intuitive process of model building into a more systematic approach, and enhancing communication ‘within a heterogeneous group. McLean and Shepherd (69, p. Sl] point out that SM “compels open discussion of the crucial relationships relating to understanding. ..the struc~ ture of the system.” SM fits the notion of modeling expressed by Holling and Goldberg: ‘We would rather be roughly right about a whole system than precisely right about a trivial part of that system... model that ean be probed and explored in a simulated ‘world... becomes an evolving device of sel/-instruction. Its ‘value is not so much to give answers as to generate better questions; not to define policy but to expose some of the ‘consequences of alternative policies (42, pp. 22. 24] ‘The following are alternate descriptions of structural mod- ling John Warfield: ‘Structural modeling is a methodology which “employs waphics and words in carefully defined patterns (0 1y the structure of a complex issue, a system, OF & field of study” (101} (0018-9472 /80/1200-0807S00.75 ©1980 IEEEMick McLean and P. Shepherd: What do we mean by a structural model? The term ‘structure’ has been defined as “the way in which the component parts of the complex whole are inter-related: that which is made up of many components parts. Structure in this sense is inherent in any mathematical model; all such models consist of components with interactions specified between them. Thus the essence of structural modeling is one of emphasis only; that is to say that a structural model focuses on the task of electing the components of a model and explicitly stating the interac- tions between them [68]. Dennis Cearlock: Emphasis is placed on determining if there is a coupling between variables and the relative importance of the coupling, rather than on developing exact mathematical relationships and precisely determining the associated coefficients. As such, preprogrammed, simple functional forms are used for specifying the coupling relation be- tween variables composing a system. Thus, structural modeling is concemed with identifying tends and equi librium states rather than quantitative precision (14) ‘As an aside, there is a term in the system field which we hear as holism, but often do not take note that there are two distinct ideas involved: one spelled wholism, and the other, holism. In both cases reference is to an appreciation of the whole. The difference is in the process of gaining that appreciation. ‘The wholism process involves reducing the whole to its parts, studying them, (a process akin to what is called reductionism) then extrapolating an understanding of the whole, Given the hypothesis (accepted by most systems thinkers) that the “whole is greater than just the sum of its parts,” the extrapolation is not possible without somehow considering the whole as something transcending the as- sortment of parts and interactions. Thus, after identifying the components of the whole and identifying their inter- connections, the resulting model is studied as a whole to develop a Gestalt appreciation. Holism, on the other hand, is an approach wherein attention is focused directly on the whole and its char- acteristics as a whole, without any recourse to reduction.’ ‘An example of holism is the process an architect goes through in developing a concept of a building when con- sidering it from its stylistic aspects. After conception of the style (a holistic process), the architect considers the various components of the building (electrical, mechani- cal, structural, etc.) and their “interconnettions,” and evolves a whole which embodies his original concept (the latter is a wholistic process). Structural modeling, then, can be viewed as a wholistic process in that the user aspires to gain an overall appreci- ation of the system as a whole by studying a structural ‘model of the elements which comprise the system. pa tele) Trea uae two ages tne appaeny Ape bask td fork “The usage in tis paper i after [83},[66} a0d (87) See ene rena eee eae ETE [EEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 4C-10, NO. 12, DECEXRER 1960 As stated above a structural model is a collection of elements and their relationships. The SM tools dealt with in this paper assume that the set of elements is given, and focus on the investigation of relationships among them, viewed from the standpoint of structure. Accordingly it was decided during the course of the study upon which this paper is based to call these methods structuring tools. It was felt however that defining the elements in the first lace was a critical part of the SM process as well. Since this aspect is not usually addressed in the application of tools used for structuring alone, it was decided to call explicit attention to this prerequisite phase, and give it a name of its own: the generating phase. In addition an effort was made to seek out methods that would assist in this equally vital part of the process. The “tools” so inventoried, conveniently dubbed “generating tools” (sim- ilar terminology appears in [17], [24], and (102)) are dis- ‘cussed in a separate paper [56]? The structuring tools evaluated in the present study consider only pairwise relationships among the elements. There is, however, research underway which considers more general relationships (e.g., see (8), [12], (14), [49], {50}, and {51)). In the context of considering all possible interactions among a group of elements, the pairwise constraint is rather severe. Recognizing this constraint, some structuring tools provide extended capabilities auxiliary” relationships, delays, rate relationships) for handling certain nonpairwise connections. Even given such extensions, though, it should be kept clearly in mind that SM tools, as presently developed, assist the participant in exploring only a small subset of possible connections among 2 set of elements. Still, even though strictly pair- wise in capability, structural modeling is often worth undertaking if only for creating an initial structure on which further investigation can be based. The remainder of this paper is organized as follows. Section II briefly describes the principal works on SM and presents a glossary of 39 selected tools, with eross refer- ences to principal authors or developers. Section III pre- sents some comparative information concerning the SM structuring tools evaluated during the course of this study. Section IV develops a “Map” which portrays in an organized fashion the relationships between 1) the basics of digraph theory; 2) the extended assumptions afd data requirements for certain of the SM tools; and 3). the types of questions that can be asked (or infor- ‘mation inferred) about the systems being modeled. ‘Section V gives some practical recommendations concern- ing the use of the SM tools evaluated, and some sugges- tions for future developmental work in this area. The last section lists the references. The entries are organized alphabetically for ease in looking up author citations ‘given in Table I. 2 For the person interested in orchestrating (faciitating the SM pro- cess, there are 2 number of important aspects of which to be aware. ‘Some of thee are discusted in (56)Levpans: STRUCTURAL HoDELING Il Brier Descriptions oF PRINCIPAL Works A. Theoretical Foundations ‘Two foundational works in the development of struc- tural modeling appeared in the mid-1960's: 1) Notes on the ‘Synthesis of Form, Alexander (1; and 2) Structural Mod- els: An Introduction 10 the Theory of Directed Graphs, Harary, Norman, and Cartwright (38}. ‘Alexander sets forth a general account of the nature of design problems: a design problem is one of formulating a structure appropriate to a specific context. He outlines a method beginning with a large list of elements and uses mathematical analysis to develop diagrams which relate the elements. His mathematical analysis has been given considerable attention—a fact somewhat disconcerting to Alexander, who feels that the importance lies in the dia- grams themselves, not the method used to create them. C._L. Owen has developed a clustering algorithm, DCMPOS, based on Alexander's work [77]. This algo- rithm has recently been used by Haruna and Komoda Gapan) [39] for structural sensitivity analysis. Alexander's method emphasizes the formation and/or identification of a meaningful decomposition of the set of elements by considering the connections between them. ‘The set is partitioned in such a way that connections between subsets are of minimal importance. A measure of, the amount of “information” crossing the boundaries of a particular partition is developed. A heuristic technique is, used to search the space of possible partitions until the amount of information crossing the boundaries is mini- mized. In later work Alexander discovered that such an elaborate scheme was not necessary; subsets can be devel- oped naturally in a piecemeal fashion. ‘Alexander employed a concept he called the variable,” based on the premise that the process of design, analysis, or problem solving consists of achieving fitness between the form and the context. The “context” defines the problem; the “form” is a.solution. The misfit variable process, then, is one of identifying “misfits,” then proceed ing to eliminate them. ‘This misfit notion derives its strength from Alexander's observation that “...when we speak of a bad fit, we'refer to a single identifiable property of an ensemble, which is immediate in experience. Whenever an instance of misfit ‘occurs in an ensemble, we are able to point specifically at what fails and to describe it.” In describing a good fit, on the other hand, “...reference must be made to the simul- taneous satisfaction of a number of requirements” (I, p. 23}. While the two assertions are logically the same, the former process-is in reality much more straightforward in implementation, thus offering a good operating proce- dure—at least as a starting point in attacking complex problems, Harary, Norman, and Cartwright discuss at length the basic principles and mathematical properties of directed graphs (digraphs). In their view a digraph is the funda- mental structure diagram. The various equivalencies be- ‘tween digraphs and matrices are developed, and the con- Berea renee teen eTOCs See cepts of adjacency, reachability, and connectedness are discussed. The notion of “distance” and the effects of removal of points or lines, all of which have strong impli- cations for the “criticality” of elements and connections, are studied in detail. Several special cases and extensions of digraphs are then analyzed. Building on these earlier works, Warfield has made numerous contributions {100}, (101, (102}. His principal work in this context is the development of ISM (see also 57] and (63, vol. 2, appendix BD), a method for forming hierarchical structures for a set of elements, based on a minimum of information about the pairwise connections. Warfield recognized that, as the number of elements in- creases, the consideration of all possible connections be- comes very tedious. By assuming transitivity (and possibly asymmetry and irreflexivity) of interconnections, one can infer most of the interconnections from knowledge of just a few of the connections. Thus, to reduce the tedium, a computer can do the bookkeeping and make the in- ferences. The computer asks users to enter pairwise con- nections one at a time, until enough information to con- struct a reachability matrix has been obtained. The re- achability matrix is converted to block-triangular form, which illuminates the multilevel hierarchical form. Warfield emphasizes the importance of abstracting em- pirical observations into a content-free model which can be easily manipulated. The results of the manipulation must then be interpreted in terms of empirical informa- tion; hence the name interpretive structural modeling (isM). Besides the ISM methodology, Warfield makes numer- ous other contributions in the area of digraph develop- ment. He suggests when it may or may not be appropriate to assume transitivity and what can be done if transitivity does not obtain. For example, it may be easier to develop a nondirected graph before developing a digraph. Warfield also explores the problem of developing a useful digraph, given its matrix equivalent. ‘Another author who has made important contributions to the development of structural models is Roberts (81]- His work deals primarily with the development and analy- sis of signed and weighted digraphs. A group of experts (or an individual) is used to establish a preliminary list of elements and then select the most important. The experts also determine which elements are connected, and specify the sign and strength of the connections. Roberts has developed stability criteria for evaluating signed and weighted digraphs in terms of the eigenvalues of their corresponding characteristic matrices. Thesé criteria em ploy a model of the dynamics of a digraph called the “pulse process model.” The pulse process model can be applied repeatedly to determine the stability of the struc- ture without computing the eigenvalues. A computer pro- gram called GENSIM [2] has been developed for this purpose. Most recently Cearlock (13] has completed a critical evaluation of digraphs, pulse analysis, loop analysis, ma- trix powering, XIMP, Impact, KSIM2, and QSIM. He focuses on the theoretical limitations of these tools. In his10 view, weighted digraphs, pulse analysis, Impact, and ma- trix powering are analogous to linear static input-output models, just as loop analysis, KSIM2, and XIMP are analogous to linear dynamic input-output models. QSIM, he suggests, is appropriate for nonlinear systems. B, Surveys Delp ef al. [17] have recently prepared a survey of planning tools. Among the 43 tools succinctly described are many structural modeling techniques. Baldwin (5) presents 21 applications-oriented papers relying heavily fon ISM, Other useful surveys are those of the Stanford Research Institute (74] and the San Jose State College ‘Cybernetic Systems Program [88], . Specific Methods J) INTERORD: Mf the conditions of transitivity and asymmetry obtain, and connectivity can be assumed as well, then the elements can be placed into a simple order. Kehoe and Cliff [48] have developed an algorithm for ordering a list of elements by means of a minimum ‘number of paired comparisons. The basic process is to ask for these comparisons one at a time, using matrix powering? after each question to determine if a complete order is implied. If not, two indices are computed and used to determine the best comparison to ask next. The purpose of this strategy is to have a computer perform a large number of comparisons thereby reducing the num- ber a person must make. A computer program called INTERORD has been developed [48] to carry out the strategy. 2) ELECTRE: ELECTRE is a French development {83} and [84)) and the basic literature is in French, The ‘most concise introduction is found in [63, vol. 2, appendix C}. The person performing the technician role in a group session [56] may want to refer to Tawfik [92] for more theoretical background if French is beyond his/her re- sources. Tawfik provides a listing of the program. ELECTRE is a computerized algorithm for ranking a set of alternatives, when each has been rated on several criteria, or by several evaluators. The ELECTRE algo- rithm avoids some of the problems inherent in conven- tional multiplicative evaluation techniques. Comparisons among the alternatives (one pair at a time) are made for each criterion, and an index is calculated. The index is tested against three thresholds (Which are set by the user) and the relationship between each pair of alternatives is determined as “strongly preferred to,” “ weakly preferred 10,” or “no preference.” A ranking is developed from the preference matrix. 3) GAF: The development of structuring type tools (ST) has taken a somewhat different course in France. One leading author is Gerardin (30). [31], who describes a tool called “Guidelines for Alternative Futures” (GAF). GAF involves the development of a cross-impact matrix incor orating both qualitative and quantitative entries. Raising x matrix to an integer power, {EEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. $4C-10, NO. 12, DECHAER 1980 ‘The quantitative form is an adjacency matrix which can be manipulated to obtain a “leveled-graph equivalent,” which helps identify the sources and sinks. Even more important, according to Gerardin, is the identification of ‘major loops in the digraph of the matrix. This is done using matrix powering to determine the strength of each Joop. GAF also uses a matrix of time delays to determine the time lags associated with each loop. By inspecting the strength, sign, and time lag of the various loops, one can determine how to regulate or modify the behavior of the system effectively. Gerardin has commented on the im- portance of subsystem identification but does not provide method for accomplishing it. He has proposed, however, an extension of the structural matrix wherein the matrix ‘entries are essentially state transition diagrams. This extension apparently bas not yet been implemented. Gerardin has been doing research on cross-impact analy- sis in parallel with his work on structuring tools. 4) SMICT4, MICMAC/MICROBE: Also in France, Duperrin and Godet have developed structural models for use in conjunction with crossimpact model of their design, SMIC74 [19], [73]. One method, referred to as MICMAC [20}, is used to rank a set of elements in terms ‘of importance. Importance is measured in a fashion simi lar to that of GAF—by raising a signed adjacency matrix to a power equal to the number of elements. This provides an indicator of the number of indirect “chains of in- fluence” in which each element is involved. This indicator is considered along with the number of direct relationships in which the element is involved. The “prime” variables are selected, and the reduced matrix serves as input to the MICROBE method. MICROBE performs a detailed anal- ysis of the various loops in which each element is in- volved. Duperrin and Godet’s conjectual systems analysis (CSA) is essentially MICMAC plus MICROBE plus SMIC74, 5) Spin: Building on the work in France and that of Roberts and others, McLean er af. developed a com- puterized structuring-type tool called Spin. The basic the- retical document is McLean, Shepherd, and Curnow (71]. In addition to allowing users to perform pulse analysis, matrix powering, and feedback loop analysis, Spin also permits clustering, clique analysis, ordering, and simplex analysis. These analyses are useful for identifying subsys- tems and discovering patterns of relationship within a Weighted digraph, and will be summarized shortly. McLean has made an important contribution by organiz- ing various methods into a common framework. This permits structural analysis to become part of the structure development process. For McLean, the role of structur- ing-type tools is to help “get the problem right” by permitting consideration of multiple models during prob- lem formulation, that is, by helping to integrate the gener- ating and structuring phases. ‘Clustering involves converting the adjacency matrix to a “distance matrix,” upon which statistical cluster analysis is then performed. McLean uses a basic single-linkage method to aggregate elements into clusters according to their relative distances. The two closest elements areLENDARS: STRUCTURAL MODELING lumped together first and the new distance matrix is formed. The process is then repeated successively until one large cluster results. The results of all the stages are presented as a “dendogram” (71, pp. 46~49] which can be inspected to determine the subsystems. A clique of a digraph is a subgraph for which all of the elements are connected. Cliques can be considered as subsystems, and for some purposes can be treated as a single element. Cliques can be easily identified from an unsigned, unweighted, nondirected graph. A digraph is ‘easily converted to this form. Ordering refers to the use of elementary row and col- umn operations to reorder the adjacency matrix in order to achieve a nearly block-diagonal form. This process is similar to that used in ISM to obtain a block triangular form, except that Spin operates on a weighted adjacency ‘matrix, rather than a threshold reachability matrix. Once the adjacency matrix is in a nearly block diagonal form, a process called “density-ratio maximization” is carried out. For each potential subsystem, a coefficient of isolation is computed. The best subsystems are then presented by means of a pseudo-dendogram. Simplex analysis is another method for grouping ele- ‘ments that have a common property (eg, they have similar interconnections with other elements). For exam- ple, if elements A and B are both influenced by elements E and H only, then A and B form a simplex. ‘The user’s guide to Spin commands is the SPIv-DoC file included in the Spin software tapes (89). 6) Loop Analysis: Working independently, Levins (60] has developed a method for assessing the stability of weighted digraph structures. His work is similar to that of Roberts, except that Levins’ criteria are a translation of the Routh—Hurwitz stability criteria for systems of linear differential equations. The translation allows the stability to be evaluated by analyzing the loops of the digraph rather than its eigenvalues. For small digraphs, a com Puter is not needed for use of Levins’ eriteria. 7) KSIM: KSIM (Kane (44] and Kruzic (53), a well known SM tool, differs somewhat from those aforemen- tioned in that it focuses on the dynamic properties of a structure. In the KSIM process, an expert group is first asked to develop an interaction matrix (or weighted di- gzaph) and to determine the initial values of the elements. This information is entered into a computer program which projects the values of the elements over time, taking all of the interconnections into account. The underlying model assumes that: I) the system variables (elements) are bounded; 2) variables increase or decrease depending on the net impact of all the other variables; 3) a variable's response to ifapacts decreases as the variable approaches its bounds; 4) a variable's influence is proportional to its value; and 5) complex interactions are described by a looped network of pairwise interactions. A recursive for- mula based on these assumptions js applied to the initial values. The variables follow a basically S-shaped trajec- tory toward the upper or lower limit. ‘Advanced versions are available. KSIM2 [45] allows the use of weights dependent upon the values of the elements, = ‘and uses a double-weighting scheme. The latter is an attempt to enrich the model by accounting explicitly for the impact of rapid changes in the values of the elements (rate impacts). In KSIM3 [53], the user has a choice of ten functional forms for describing the coupling of any two variables. This version, then, is similar to QSIM in flexibil- ity. KSIM computer programs are documented in a very useful manual by Kruzic and Sandys (54). The person filling the technician role [56] may wish to consult Cearlock (13}, Suta (91), and Burns (9] for a discussion of limitations. ‘The implementation at Portland State University of the code originated at Stanford Research Institute has eo- hanced the conversational and interactive features of KSIM. 8) QSIM: A similar but somewhat more elaborate tool known as QSIM was developed independently by Wakeland [96]. In QSIM, variables are not automatically bounded; nor need the coefficients be constant, or the equations linear. Users are first asked to develop an unsigned unweighted adjacency matrix for the elements. Then each binary connection is elaborated in terms of @ graph called an “interaction plot,” portraying the rate of change of one element as a function of another. Each such plot is treated as one term in a set of first-order differential equations, which are then numerically integrated. To model complex interconnections among ‘more than two elements, QSIM allows for the use of “auxiliary variables.” These are functions of the other elements of the model, and may be polynomial, piecewise- linear, or tabulated. The computer program is docu- ‘mented in a user's guide [98]. This guide contains a “walk through” of the QSIM modeling process. 9) Impact: Another structuring tool which computes the time behavior of a weighted digraph model is Impact {76}. Impact is essentially a very specific type of pulse process called an autonomous, continuous pulse process (Gee Roberts (81) with the additional charcteristic that variables are bounded. Hence, it appears to operate much like KSIM, though the underlying mathematical models ‘are quite different. Users develop a weighted digraph which is entered into the computer in order to estimate behavior and test alternatives. The Impact guidebook contains not only a description of the modeling process, but also a Fortran listing of the relatively short Impact Program. This guidebook has been designed as a class- ‘oom tutorial. 10) XIMP: Moll aad Woodside [75] haye developed a computerized structuring tool called XIMP which is based ‘on a slightly modified version of KSIM. XIMP features a user-interactive computer implementation that has builtin structural analysis capabilities. The principal modification from KSIM is the incorporation of “base” values. The impact of an element at a point in time is proportional to the difference between its base value and its value at that time. Ifthe element has a value less than its base, then the sign of its impact is reversed. ‘The structural analysis of XIMP includes stability anal- ysis, sensitivity analysis, parameter identification, andam tracking optimization. The stability analysis is closely related to that of Roberts. Criteria are developed by ing the XIMP equations near a stationary point and investigating the eigenvalues. Stationary points occur when each ty analysis in XIMP is performed by com- puting the rate of change of system elements with respect to the entires in the adjacency matrix. Two types of indices are used. The first type is the sensitivity of one element to one matrix entry. The second is the sensitivity fof the overall model with respect to a specific matrix, ‘entry. The indices are computed by perturbing the value of the matrix entry of a small amount, then summing over time the changes in the values of the elements. The sum is divided by the size of the perturbation to obtain a relative Parameter identification and tracking optimization are variations of the same basic process—automated adjust- ‘ment of matrix entries in order to achieve desired trajecto- ries of the element values over time. Though this process may fall outside the definition of SM, we touch on it since it is incorporated in the XIMP computer program. Moll and Woodside accomplish the automated parameter adjustment by formulating it as a mathematical optimiza- tion problem which is solved using the Marquardt heuris- tic search techniques. 11) ISM: Interpretive structural modeling (ISM) [100], (102), is a computerized algorithm for arranging a set of elements in an ordered sequence or hierarchy in accor- dance with a given ordering relationship. By assuming the connections are transitive and pairwise, the computer ‘obviates the need to compare all possible pairs. It requests the minimum number of comparisons, uses the algorithm to calculate the temainder and quickly determines the overall ranking of the elements. While the procedure seems unnecessary if the number of elements is small, itis very useful for a large set. Warfield, the foremost proponent of ISM, emphasizes the importance of abstracting empirical observations into content-free model which can be easily manipulated. ‘The results must then be interpreted in terms of empirical information; hence, the name interpretive structural mod- cling. If theie are a number of elements which can all be reached from one another, the first portion of the ISM program will coalesce all of them into a cycle and subse- quently treat the cycle as a single clement. A “cycle resolution” process may subsequently be used. In this process, the participants provide additional information about the connections among the elements in the cycle by filling in a weighted matrix. Each matrix entry represents the “intensity” of a connection in the cycle; the program then uses a threshold concept to determine a suitable binary matrix in a series of steps. Questions which have been raised concerning ISM in- clude the “first element syndrome,” (i. does the first clement examined bias later comparisons?) and the voter paradox (Je, individual transitivity but group intran: ity), Neither appears to present serious constraints, [EE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SHC-10, NO. 12, DECEAOEX 1980 D. Glossary of Selected Tools Table I is a glossary of a number of the tools consid- cred in the aforementioned survey, along with references to the principal developer /author of the tool. The bibliog- raphy to this paper is alphabetized, so that references to these developer/authors are easily accessible, The right- ‘most coluran on the table contains some short descriptives or other comments. TIL, Comparison oF Nive SELECTED ‘StrucTURING-TyPE Toots A. Selection of the Tools for Evaluation At the outset of the survey on which this paper is based, more than 100 candidate methods were identified, based ‘on an initially broad definition of what structural model ing was to include. Criteria were developed to reduce this toa more manageable number. Approximately 25 percent were set aside as generating-type tools, and approximately 40 percent eliminated for various other reasons: lack of documentation, lack of generality, quantitative as opposed to geometric emphasis, ete. ‘This process of elimination left us with those SM tools that assist in organization of or analysis of interconnect tions identified as relevant to a specific context (see Table 1), Once the geometrie/graphic representation is availa- bie, various properties are determined, either by manual means (tracing effects through the graph) or by entering information into a computer concerning the nature of the connections (as their sign, magnitude, and/or functional form). ‘We then applied the following practical criteria to select the tools for in-depth study. 1) Each is fully implemented and available for use. 2) Each is readily understood and used by persons skilled in mathematics and modeling. Other persons ‘or groups must be readily able to understand and use the method under the guidance of skilled facili- tators and technicians (56). 3) Each is general in applicability. For example, a computer algorithm used strictly to manipulate sgzaph structures for easier computer storage is too specific. . 4) Each permits the use of subjective data In some cases two or more tools incorporated the same algorithm. In each such case, only the more comprehen- sive tool was evaluated (but only after verification that the algorithms were identical). ‘One point is to be stressed: while seven of the nine tools chosen are implemented on a computer, structural model- ing is not inexorably linked to computers; much effective analysis can be done using pencil and paper. B. The Comparisons Nine specific methods involving digraphs and/or matrices were selected for detailed evaluation. Each of these tools can be applied in a few days at relatively small{LENDanIS: STRUCTURAL MODELING TABLEL Guossany oF Seiecren Toots Nenear ‘Devoperor Anbar on hates Dae DALIXINe —Twott 1995 DASSESS— Wateand 70 no Deperia 19050 perc: Unekty te DOCMPOS Owe” 90 DEINE Conon Me DDiewphs —Hareyeatal ans Rees tone oreame are tn S)RLECTREM Roy ay & Reser im ELencxme Dukey IDEDUMP Rar MREVENTX —Bloow CAF Gate to cemin Allee 19Gb Pale 1oGnrs Hades Marcy s9)ssa farms eed do syameee Aion Spee Forte =m syrosan Foon, ws teste. a cost. Each, however, requires specialists if proper use is to be assured. Some of the tools are very versatile because the elements and connections can be used to carry almost any conceivable meaning; others admit fewer meanings and are hence more limited in application. ‘A number of characteristics were deemed useful for ‘comparison of the various tools. These characteristics are grouped into five categories: 1) basic characteristics, 2) connection-related characteristics, 3) algorithm properties, 4) group aspects, 5) computer implementation properties. ‘The comparisons based on categories 1) through 4) are collected in Table Il; those based on category 5) are collected in Table III. In Fig. 1 the tools ate organized according to one of the characteristics of category 3). Fig, 2 is a graphic representation comparing the seven com- puterized methods as a composite of several of the char- acteristics. The following are the specific tool characteristics within each of the categories 1) Basie Characteristics: Number of elements: The number of elements that can be effectively handled varies considerably. Assign- ‘ment of an upper or lower limit is highly subjective. We offer our estimates only as guidelines for newcomers to ‘SM; the experienced modeler will develop his/her own relative to specific tools. Computer requirement: Most methods require a com- puter; however, one of the methods can be run on a Programmable calculator, and several achieve useful re- sults by manual tracing of implications. 2) Connection-Related Characteristics: Cumulative and proportional: Assuming a connection from variable x, to variable x,, if x, changes even though remains constant, then the interaction is cumulative; if x; changes only when x, changes, then the interaction is Proportional (10}, [18]. More complex relationships are possible, and some methods permit the following. Functionally dependent connection: Relation between ‘two variables is nonconstant and dependent on other factors. Example: the interest rate operating on a savings ‘account is five percent for first five years and six percent thereafter. Delayed connection: The impact may take effect only after a time delay. Example: the effect of snowfall on ‘runoff. Only one of the methods permits specification of time delays. Nonpairwise connection: Two or more variables act together to affect a third variable. (For example, energy ut into achieving a goal depends on the amount of work remaining.) In general, a coupling is nonpairwise if the impacts cannot be separated into terms containing only one variable each. 3) Algorithm Properties: Essential assumptions: A number of methods assume linearity; one assumes transitivity. Only some methodscry [IEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-I0, NO, 12, DECHRER 1980 TABLET (Coreanssot or Nove Seizcren Srmucrunmvo-Tyv® Tools ACCORDMO TO CATEGORY I THROUOH 4 CHARACTERSSTICS TREES GM ELECTRE NETWORK IMPACT SPIN —XIMP_——KSIMN SIME ‘avin mer of oe eo) oe ” « ” nr) Compa eit? No ee we weve Proprio NAMA Ate we Ne ve . Feely Noo ee Ne Ne eo pact Beto i me Ye Ne Ne Ne Via aia pein Ouiative oo ve ve ve 1 nora cea ep emphasis Mee we te Ne Me atone Sones ‘Symmes agneme ed teat Sindh Sime rep pct Expt aowace Ne Yet eto Ne Ne Ne Ne Ne ‘sung ey ay ny ny Ey Mertens Moderne Hd [Lelogistie; NA= not applicable; *»= Waller alternative (9) “In specific applications trees and networks may have extremely large numbers of elements; however, when uted ‘heuristically to organize information, they quickly become unmanageable as the umber of elements increases can incorporate empirical data when it is available (how- ever, see Gur (36). (Qualitative emphasis: The focus is on inexact, subjec- tive input producing meaningful output (but not precise, quantitative forecasts of behavior). | Causal lop emphasis: The focus is on tracing causal om ETRE paths and discovery of feedback loops, if any. This pro- cess may pertain either in the building of the model or in cater] the analysis stage. Time behavior (see Fig. 1): Four descriptives are used ee for this subeatezory: ah static—constant over time transient—short-term response to change dynamic, uncalibrated—algorithm computes behavior over successive time incre- cement aa ments but an increment has antine ahe y om a no precisely defined interpre- so fected Sea tation; the latter is left to the pele user. ie Speman dynamic, calibrated algorithm , computes behavior toni ayo ‘over successive time incre- ‘cu ments; an increment has a Fig 1, Organisation of nine selected structuring-type tools according specific temporal definition. to aie and dynamic analysis capabilites,‘Lewpans: sraucTURAL MODELING a —_ rat a d Aa mie ts se Fig 2 Comparison of seven selected computerized strcturng-ype tools according 19 & componite of the criteria in ‘eategories 1) through 4). TABLE It ‘Covernniso oF Seven ConruTenszzn STRUCTURINO-TYPE TOOLS ACCORDING TO HUMAN-MACiIRE INTEREACE ‘Gusacrmusnics GM GLECTRE SIN wcT. NIMS 2 Prods copue ing rer : som = = > 3 Syn a ee cee eo een 5: Omran = a ey =e 5. onde ne ching 2 & - & SB SF = 9 Rapdtepom rn ee) 10, ay omy moet em eee 1 Bayi a ey = = = i tywaeat moe eee "Fite yersion: yes 4 Group Aspects: Explicit allowance for disagreement and for multiple criteria: One tool in particular makes this allowance. Explicitly designed group process for making subjective Judgements: Two of the tools incorporate this feature. Ease of use and communicability: This entry attempts a summary measure of the number of judgments required, the mathematical complexity, type of data required, form of output, and sophistication of human-machine interface (Gescriptives used: easy, moderate, hard). 5) Computer Implementation—Criteria Dealing with Hu- ‘man-Machine Interface: Anticipates user errors: Catches errors without losing data; provides useful defaults. Provides complete editing diagnostics: Tells exactly what went wrong if an error occurs so that it can be easily fixed. Provides display upon user demand: User controls what is printed at the terminal. ‘Multiple commands: Some purpose can be achieved. Operator inputs echoed: Traces of inputs readily avail- able (other than original). Easy to learn and use. Provides a tutor for self-teaching: Either a “long prompt” level or a special procedure. Versatile: Easy to add new commands; modular de- sign, command processor oriented. Rapid response: Minute turn-around times. Easy to expand model: Add variables. Records inputs for succeeding runs: Ability to stop partway through the procedure without losing data. Easy to save model: By means of a simple command, It must Be kept in mind, especially with respect to the category 5) characteristics (tabulated in Table III) that the comparisons made are subject to further developments. Judgments made were based on the computer implemen- tations available at the time of the survey. A potential user should seek the most up-to-date version of the tool ‘when embarking upon its use. A difficulty involved in all of the above is that methods (as opposed to specific models) are being compared. Al- though procedures and criteria for evaluating models abound in the literature (e.g. (79, pp. 12-48), the above criteria had to be specifically developed for this particularEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. $%4C-10, NO. 12, DECELOER 1980 816 WOLLWWwOSNI ‘4907 wowa0a34 wonesaudianuy J T ws) Finny door. J Awenny 1 wolyiuosNt ‘ALIALLISNAS. 1X31N09 09:3 / wainoud aa00% ANouvuaIH GASTILIAN[Note: pgs. 816, 818 & 819 have intentionally been moved to follow pg. 834 J evaluation. An important additional dimension to keep in mind when comparing any of the tools is the skillfulness of the user. The best of tools in unskilled hands are likely to produce results inferior to those of “lesser” tools in skilled hands. Finally it should be remembered that SM is often a group activity. The seven selected computer-implemented tools are particularly useful for such applications. As such, their use involves three distinet roles: the method techni- cian, the facilitator, and the participants. These roles are discussed in [56] IV, Tue “Map” Since the various structuring tools have a common mathematical basis, it is possible to organize them accord- ing to which specific aspects of the theory (of digraphs) ‘each is based upon, and further, upon what specific addi tional assumptions and/or data each one requires for its, use. In addition since each of the structuring tools leads to the exposition of certain kinds of data for a given system, and each allows for the participants to build up their insight and understanding according to different qualities of the system being studied, these aspects too can be used for describing and distinguishing the various structuring tools. NoTaTiON Make assumption ‘Aad information Encicle roman identities output. Enciced arabic number —_—— a, a7 It was decided to build a structural model of the theory ‘of digraphs. to then add in representations of selected SM tools, and finally, to incorporate reference to the kinds of information that can be inferred about the system being studied. The result is what we call the “Map.” By itself, this Map is exceedingly complex (as can be seen in Fig. 3). Accordingly an explanatory method is incorporated here to lead the reader through the develop- ment of the Map section-by-section. Other presentations of the same information follow the exposition of the Map. ‘These presentations, one for each possible variety of out- Put, constitute Fig. 4. More about this later. ‘The Map in itself is a structural model. As such, it consists of “elements” and “connections.” The “elements” are states of knowledge, either intermediate or final. The “connections” are processes of refinement (called steps) which lead us from one state to the next. ‘Three kinds of activity may occur in a step: 1) an additional assumption is made (A), 2) further information is added (3), and/or 3) an algorithm or operation is executed («). ‘The notation to be used is as follows. oureut Execute algorithm or operation bet (and indicates “eitheror” choice-— ® [EeE TRANSACTIONS OW SYSTINS, MAN, AND CYBERNETICS, VOL. SUC-I0, NO. 12, DECEDOEX 1980 @ ‘none context ‘Ati Peeeecton, EBaeerion @ SSRiCchara PROB conte cay gee i Bou OE Ratings Compson Asm Eoeczine INARY INCIDENCE, IATRDL/OIGRAPH ‘ester earn Zz, Shee ‘oo ‘SGNED mclaEner, iktandbioRarn T hoe ie Mes WRGHTEG scneD, theraence MATRIbicRAPH fon Suesvsrew © ienrirreation ELEMENT Pius PROBLEM CONTEXT @) ELEMENTS Plus raoaten conTeRT | Pan Powe 7 Pittien 5 eee Set, roman CES Bam | ERT oa “Ehane Pt Woah Number oe WEIGHTED, SIGNED, E bane Vatues Foie Eee ee SIDENCE MATATCDICR ele set fies pe vim Boy PaEranTe8, SEN. eience MTRICBIGRAn fede foe ern fon]_{oAl TAITICAL. (Gounces Eteaeers ao Swolratis Ss. ier Ue eet FEEpEAEE LOOP invORMaTIOR Emerge Fig. 4. Summary Now charts taken from SM Map of Fig. 3 (read top to ‘bottom.{ENDARIS: STRUCTURAL MODELING © ean ‘ASymmane be EOvections TARY EIOENTE, BRERICOIGRAPH Eee Ficus mcsoence, fuatnneoranana’ | Tara wie Sten meee, washus [tora al 2am aera) [ELEMENTS PLUS PAOBLEW CONTEXT @) Tenens os raosten conten hie Berean, ‘A Spmmune te £ Oe EP Pac Woh Es 5 ftv Waa Tene renee Marainvoionsrn’ Sia wae igeataesl Lith inti Lape Pc INEORUATION EEDeAcE LOOP Mee ‘ame Perucbtions Fensiriire iteration! TLeuenrs Pius PROBLEM CONTEXT GD] $B, mule Wht, ita tan ‘Speman oP Bisco ane ienal Vat imraT FOR Sta wane @) "ELEMENTS FLUE PROBLEM CONTERT fF Eee tna ve, ate Bama aut inrat Fon est woe @) _E hasary aoc (nae pan AD TaPaT FOR, 2 MODE yi oe Fes vce Teneo weiaeace waTnnrrenara | | comics woh tte Epona ih t Wak Wee. © toga 7 mer PET FoR sin MDE Shek tito face! [eaicnareo time Taasecrones] [fircatienareo te ranecrones] Fig 4 Contimed. @el in0 {EEE TRANSACTIONS ON SYSTIBS, MAN, AND CYBERNETICS, VOL. SMC-10, NO. 12, DECENEER 1980 In the discussion that follows, the basic text is placed on the left side of the page, with concurrent graphical rendition and explanatory remarks on the right. Following the discussion for each step, the appropriate portion of Figure 3 is reproduced in a "box," like this. @ As we are concemed with structuring tools, we as- sume there exists a list of elements (at least a tenta- tive one) and a problem context. We are in “state 1.” @® We now develop a graph or matrix. element A cement & ‘A: Assume that only pairwise relationships obtain ok between elements. oalation ” \ T not QK, oN E: Provide connection information by drawing a graph showing connections between elements. Optionally this information may be encoded in a binary connection matrix (1 indicates connection, blank indicates no connection). This representation is mathematically equivalent to the graph. Connections may have various connotations. In ordering persons by height, for example, the connec- tion could represent “is taller than.” (For this partic~ Ular relation, direction is important. See next step.)LENDARIS: STRUCTURAL HODSLING a @® For any connections, we must either A B svanemuc wens So, E: Add specific direction information, ie, pA eee Tt 8] 7 ‘A: Assume that the relation is symmetric or arrowheads. Since both directional relations (A. to B and B to A) must now be specified, the binary connection matrix of state @ is replaced c : by a square matrix. (Continuing the previous D example, “A to B” means A is taller than B.) BINARY. INCIDENCE (iarArad MATRIX GM) on the right, the indirect connection between A and Cis apparent and there is litle likelihood of misin- ferring a direct connection A to C. In the equivalent matrix, the existence of an indirect path is not so obvious. In filling out the matrix without the aid of the digraph, a 1 might have been placed in the upper Fight box. “An interaction matrix helps avoid errors of omission while the causal diagram eliminates er- a rors of commission” (Burns {9). Clearly, indirect connections are more readily recog- nized ins digaph than to's matnc tn he dian A Be ee > @ &: Add sign information. The sign must be specifi- cally defined. The symbol plus, for instance, ‘might mean “if A increases, then B increases,” or “if the rate of change of A increases, then the rate of change of B increases.” Whatever the relation, it is assumed to be monotonic over the interval in question. +]o tf SIANED DIGRAPH SIGNED IM Q-OQ @©@ © %: Add weight information for the pulse process. 45 PAB ED Example: if a ten perceat increase in the initial s 4 level of A implies a five percent increase in the a initial level of D, the weight is 0.5 (5-10). We =¢ now have a weighted, signed digraph suitable 7 for pulse processing analysis. WelGHTED SIGNED waugtteD, epg iano TMm [EEE TRANSACTIONS ON SYSTENS, MAN, AND CYBERNETICS, VOL. SMC-I0, NO. 12, DECHICAER 1980 Q-O C-e® © Alternatively, one can Z: Add weight information to the unsigned inci- dence matrix or digraph in state @. In impact analysis, interaction can have direction and strength, where a sign might lack meaning (¢.8, traffic and energy flows.) WeladTeD, HHT AN OED unstaneD IM © From our initial state (D, we can obtain a reachabil- REACHABILITY MATRIX: ity matrix as well. This representation tells us whether ‘or not one element reaches another, either directly or through intermediate elements. ‘A: Assume that all relationships are pairwise A: Assume transitivity Add relational comparisons esr ech column element RELATIONAL COMPARISONS (example): Sis A subordinate +b 6?" no Sis B subordinate to C2" YES{LENDARIS: STRUCTURAL MODELING The same reachability matrix @ can be developed from the binary incidence matrix or digraph @) as well. An operation known as transitive closure (102] is required. Perform transitive closure (on @) Matrix @ can also be reached from the weighted, unsigned incidence matrix ©, by thresholding the entries in matrix ©, then performing transitive closure on the resulting matrix. Add value for thresholding Threshold matrix © Perform transitive closure TRANSITIVE CLOSURE: ‘The information gathered thus far from the user, and represented as above, is in a form that can be manipulated and analyzed via the now well-known body of mathe- matics known as graph theory, in conjunction with Boolean algebra (see (38). Use of this mathematics af- fords insight into various properties of the model which depend entirely upon the element connections (that is, the structure). The user input consisted of basically binary information, (es oF no, one direction or another, plus or minus), and/or simple weighting. More complex information may be introduced into the SM process, when available. 2a4 {EEE TRAKSACTIONS ON SYSTBAS, MAN, AND CYBERNETICS, VOL. SMC-10, NO. 12, DECEAEER 1980 @® At state @, instead of a simple sign or weight information, Z: Add interaction plots which relate pairs of ele- ments. (Note that the relation now need not be linear.) Further, Es Add start values for the variables and their rates 6f change ‘The results constitute input for the QSIMI model © It in addition we can develop “auxiliary relations,” the pairwise assumption (see @) can be relaxed, and ‘the input required by the QSIM2 model is at hand. 3: Add auxiliary relations (examele 2) Pe $ (fopulation) (Real GNP/Cagita) Rate of Change hs, ea Real GUP/Cagtta. A Interaction’ @ A oe mess wks i % epolation() Start Value|Start Rate of Change [Sette [lect tee of Chance] Auxiliary Relation (Example): Rate of change of GNP/capita = (constant) + (population) (capital invest- ment). Note: In this example the effects of population and capital investments upon GNP/capita cannot be sep- arated (Le., nonpairwise). @ Addition of delay information to state © yields a delay matrix. In a relationship involving “delay,” A might have an impact of, say, two'units on B, but only after, say, three units of time have passed since A initiated the change. (Note: This information is used in Spin.*) Z: Add delay information. “See Table I for plossary of the acrooyms wed in this section. ABco 2] [e5) 3] Al 2g C(t Ez] WEIGHTED , SIGNED INCIDENCE: MATRIX ABc Dp DELAY MATRIX ‘Appropciate units enceEBNDARIS: STRUCTURAL MODELING as @ Addition of cumulative weights (see Section III-B) Combined caldna and beta matrix and initial values at state @ yields input for the iti Be oo KSIMI model. (The KSIMI model uses a weighted, signed digraph or matrix: it is called the “alpha” matrix) E: Add initial values Z: Add cumulative weights. ® Addition of proportional weights (ee Section II-B) to @ yields input for the KSIM2 model. These ‘weights (similar to the pulse process weights) con- stitute a second weighted, signed digraph or matrix, called the “beta” matrix by KSIM. E: Add proportional weights According to Kane the numbers above the slanted line denote impact of one element on the rate of change of another, say, A on B. The numbers below the slanted line denote impact of one elemeat’s rate of change on that of another. Apco QNg> ® Addition of functional impact relationships among fxg Tabled hans the oariables (as opposed to the constants) in the alpha and/or beta matrices yields the input for the KSIM3 model. Note: KSIM3 allows a choice from 2 list of ten functional forms. The interaction plots used in QSIMI are similar to, one of the functional relationships in KSIM3. Zi Add selection of functional relations and associ- Functions From Data to be provided ated data which to select (fer selected function) Ny (Rio ValueHaving thus far evolved the intermediate states and their relationships, we now turn to the final (or “output”) states yielded by application of the selected SM tools to this intermediate knowledge. In notational terms we will link the “knowledge” circles (previously referred to as states) with the output symbols (C2) by means, of a variety of algorithms (+). Tt must be recognized that the distinction between inter- mediate and final states is not unambiguous. An inter- mediate state may prove in itself to be a final state, if the practitioner has derived sufficient insight from the process to this point. In other words, the learning experience gained from creating a digraph may alone ednstitute sufi- cient “output” ‘Similarly a tinal state may serve as an intermediate state for another form of output. For example, feedback loops (here designated a “final state”) also provide a basis for another final state—known as stability analysis. In general the information draw from structural mod- cls can be categorized as follows.* Scircled numbers reler to & Map notation. Ordering: + graphical representation (e.g, a digraph) + simple ordering of elements + multilevel hierarchy + subsystem identification. Analysis oia structure: + identification of critical elements and paths, sources, and sinks. + stability + feedback loops + sensitivity. Analysis via dynamics: + input needed to obtain specified time trajectories + uncalibrated time trajectories + calibrated time trajectories, IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-10, NO, 12, DECEMBER 1980, @ ® ©The Final (“Output”) States (C=) ® Simple Ordering of Elements Begin with state D A: Assume pairwise relation transitivity asymmetry of connections connectedness of all elements ‘Add comparisons for each pair of elements Perform ordering, manually or by means of INTERORD to arrive at a simply ordered list of elements. © Matievet tierarchy ‘We may begin with state D or D. > Beginning from state (D, either ‘+: Draw the hierarchy by hand if the number of, elements is small or ‘Add criteria ‘Add weights ‘Add thresholds ‘Add scales ‘Add ratings Use ELECTRE For ELECTRE: Elements are alternatives to be structured in a multi objective environment. Criteria are factors used to determine prefer- ‘ence, such as cost, risk, ease of use, etc. Weights are degrees of importance assigned to each criterion, and degrees to which cach al- ternative meets each criterion. Thresholds are specification of the “intensity of agreement and disagreement.” ‘Scales provide the framework for the ratings. Ratings are subjective measures of the extent to which an alternative satisfies the evaluation criteria, (“Painased see @) ( Porsttuitys sa@) Aaynmehy= &—$ aK. oh not OK. AN / e o ce Ok . nat ox tf ‘\ Comparisons: Aw 2? > Amb? f(a we 2 inlar to & GASex*) Muttilevel hierarchy: (retation:*oxerts influence on’) +the'diraction’of the digrach ie, battom to Rp (or fop tb battom) ete id ay ena 2 ed typically, depends on-tg ‘being portrayed— Beginning from state @, ments” algorithm of ISM. *: Thea, cither: continue to apply this algorithm over a number of cycles, gorithm of ISM. If there are no prior structures, apply ISM to the list of elements in order to develop a par- tial structure. Use the “Partitioning on Ele- ‘employ the “Iterative Bordering” al- LLL ‘A subsystem or subdigraph is here loosely defined as a grouping of nodes and connections which is self-contained to a considerable degree, minimizing the connections beyond its borders. Identification of subdigraphs represents a form of ordering achievable with several SM tools. ‘We may begin with state @, @, or ©. Beginning from state @, ‘*: Use one of three SPIN algorithms: + clustering + simplex + cliques —+ Beginning from state @, either ts Use reordering (Spin) or *: Use Alexander's algorithnr [77] or Use visual inspection + Beginning from state ©, *: Use reordering (Spin) | \ SUBSYSTEM | SUBSYSTEM 2 [EEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 946-10, NO. 12, BECHER 1980[Lmnanis: STRUCTURAL MODELIKO @B) teenifcation of ertical elements and paths, sources and sinks “Critical” is used here in the following seases: + with regard to an element—a “large number” of lines enter and leave the element, * with regard to a path—a “large number” of lines are involved in the path. ‘A “source” element has only outgoing paths; a “sink” element bas only incoming paths. We may begin with states D, @, or ©. > Beginning from state Q, either @ is arcritical"elementt @is a’source' element @is asin element ': Determine the number of connections manu- ally (MICMAC) or 1 Use transitive closure (see D) + Beginning from state @, either Perform matrix powering or Use the indirect path algorithm (Spin) (The latter process is designed specifically for iden- tifying sources and sinks.) Beginning from state @, *: Use indirect path algorithm (Spin)330 {EEE TRANSACTIONS ON SYSTEMS, DUAN, AND CYBERNETICS, VOL. S4C-10, NO, 12, DECENODER 1980 Stain The structure itself is a major determinant of sys- tem stability. Several forms of stability analysis are available using SM tools. We may begin with states ©. ©, Oo + Beginning from states @ or ©, A: Assume linearity ‘Aad pulses Apply a pulse process (Roberts, Spin) — Beginning from state ©, either Assume linearity ‘Add pulses ‘Apply a pulse process (Roberts, Spin) ‘Assume linearity Use Levin's criteria [52] or ‘Assume boundedness and linearity ‘Add continuous pulses Use Impact pulse process. genvalues yield definitive information about stability; application of pulses is just experimentation, and unless done exhaustively, may demonstrate instability, but would not prove stability. + Beginning from state @, A: Assume logistic curve 3: Add base values 2 Use XIMP stability criteria Pulse Process: Stage 2: 2u sh Dal Next, ¢ ected and B Tecees a second jotr ~and so forth. to or (2,b constants)[Lmwpanis: erRUCTURAL MODELING + Bexinsing fom sate @ 3: Add aniatepeation based on haowldge of control theory. , ‘Suggestion: Input these simple models to see how the * fF a tool treats them. Viz, if time scale is not long enough, one Unstable Se may not see the stability or instability. Control Theory Concepts GQ Feedback Loops Feedback Loops (Example) = ‘A feedback loop is a closed path from one element back onto itself. We may begin with states @, @, @, or ©. + Beginning from states @ or @, either Use visual interpretation or 2 Assume linearity Do a loop analysis (Spin) > Beginning from states © or ©, Assume Linearity Do a loop analysis (Spin) Feadback loop: 1-23-4-5°67-8" |[IEE THANEACTIONS ON SYSTENG, MAN, AND CYBERNETICS VOL. $4C-10, WO, 12, DECEMBER 1980 Senstioiy “Sensitivity” denotes + change in structures + change in pardmeters. We may begin with states @, @, or (19). > Beginning with state ®, either A: Assume linearity *: Perform matrix powering ‘A: Assume linearity Use Leontief inverse matrix [58]. + Bepioning with state LeontieF matrix : As Assume «lois euve E-A)X=¥ B Addbae valcs Appl perturbations XIMP) — Dering with sate %: Add an interpretation based on a knowledge of control theory ‘Inverse = X=GAY(]) tru Best Sd On Given a certain desired behavior of a variable over a time interval, it is possible to determine the inputs, digraph weights and initial conditions re- Quired. Begin with state @, A: Assume logistic curves ‘Add desired time trajectory ‘Add base values : Perform tracking using XIMP wsry (BD Ueatdid Time Tees A time trajectory is uncalibrated if a time scale is not provided (or is arbitrarily assigned). We may begin with states @, @), @), or ©. > Beginning with state @, 1 Perform visual inspection This alternative is particularly effective if the system is known to be quasi-static or tran- sient; that is, if only minor change occurs over the time span considered. —+ Beginning with state ((), either ‘A: Assume a logistic curve pattern 7: Merate using KSIMI A: Assume a logistic curve pattern E: Add base values : Tterate using XIMP. + Beginning with state @, [Ar Assume a logistic curve pattern Tterate using KSIM2 > Beginning with state), AA: Assume a logistic curve pattern + Tterate using KSIM3 Calibrated Time Trajectories A time trajectory is calibrated if an explicit time axis is specified. > Begin with state @, *: erate using QSIM [EEE TRANSACTIONS OW SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-I0, NO, 12, DECEARER 1980 Note: Equilibrium states may be inferred from this level of output. ‘A composite map of all the SM tools discussed in this section is represented in Fig. 3, which now constitutes a “Map” (in itself a structural model) of SM methodologies. ‘To summarize, nodes @ through @ in the Map repre- sent the basics of digraph theory; nodes ® through (@) represent the extended assumptions and data required by the selected SM tools; and, the ten “output” symbols numbered IA-IC, HA-IID, and IIA-IIIC represent ‘three categories of information that can be determined about the system being modeled. These 23 “elements” are interconnected by the myriads of lines shown in the Map. Each of the lines is annotated by the kinds of data, assumptions, and/or operations that are required to pro- ‘gress from one node to the next. Rather than searching through all possible paths start- ing from node @ to determine the best SM tool(s) to use for 2 given application, the Map would more typically be used in the opposite direction. That is, having specified the type of information desired concerning the system under consideration, one would locate the appropriate “output” nodes, and work backwards through the map to determine 1) the SM tools that could be used, 2) what assumptions (A in the Map) are required, and 3) the kinds of data required for the calculations and/or the operations required in each path selected. To facilitate this working-backwards-through-the-Map just, described, a group of summary flow charts, one for each “output” node, is presented in Fig. 4.‘eammpaaus: STRUCTURAL MODELING. V. Some RECOMMENDATIONS Rooted as it is in engineering, structural modeling first proved itself in applications to well-structured systems; it has been highly effective in purely technological contexts. Now, however, we are increasingly being called upon to effect a kind of “technology transfer” of SM to the more ill-structured systems involving not merely techno'ogy but also individuals, society, and the natural environment. Difficulties will no doubt arise in applying the SM process in such contexts, since these are significantly more com- plex. These incipient difficulties, however, should not dis- courage us, for by and large, the tools so far developed under the rubric of” structural modeling hace been useful—as evidenced by many successful applications; decisionmakers are using some of these techniques in their decisionmaking process; and, groups are using some of them to help clarify thinking, if not for making decisions. A call to action is in order. Researchers (and practi- tioners) should become more aware of the varied efforts in this direction, and take a broad approach in refining and disseminating these tools. Useful as they are, they will only be used if they are made accessible, inexpensive, easy to use and, above all, widely familiar. Much remains to be done along these lines. ‘A huge amount of information, both technical and fundamental, has already been presented so far in this Paper. We feel it is important to beg the reader's indul- gence yet a little more, and present a list of recommenda tions® offered in a practical vein, then conclude with some suggestions* for further work needed to improve SM to cope with the more illstructured systems applications cropping up in the world of “real” problems. A. Some Practical Recommendations 1) Do Not Use Computers Unless Paper and Pencil Prove Unmanageable: Attempt to construct digraphs, trees and/or hierarchies manually first. You will be surprised how much can be learned from such an effort. Computers are needed for quasi-static systems if the number of elements is large, and for dynamic systems if complex interactions among the elements are involved. However the learning that stems from computer sessions alone has often been found disappointing. This observation docs not negate the value of structural modeling; it simply stresses the importance of manual efforts at SM before engaging the computer. 2) Do-It-Yourself Structural Modeling (with Expert Sup- port) Is Nearly Always Preferable to Buying Packaged Stud- es Done by Outside Organizations: Much, if not most, of the value of SM comes from the hands-on activity—the process rather than the output. This is where the learning is most effective and where the insights are gained, The great mathematician, Richard Bellman, makes the same Point about mathematical modeling and simulation in general: “In many cases we learn most from the construc- tion of the model” (62, p. 102}. “These ae excerpod from (63, vl. 1, pp. 142-150}. een as 3) The Important Built-in Assumptions Inherent in any SM Tool Should Be Given Explicit Attention: Neither societal nor ecological systems can always be approxi- mated by linear models. Example: loop analysis applied to linearized versions of the Volterra predator-prey equa- tions leads to serious errors in stability properties (13, p. 78} Transitivity, a requirement for ISM, may not always obtain, Example: In development of energy sources, natu- ral gas affects oil, and nuclear energy affects natural gas—but nuclear energy does not affect oil (too late in time; see (64). Pairwise-only connections between elements are as- sumed in nearly all tools (notable exception: QSIM2 and KSIM3). However, the societal relations cannot always be described as pairwise (examples: vested interests in “lob- bying,” states in intemational relations). Considerable subtlety is needed in properly interpreting the system's time behavior as represented by the various SM tools. Some of the representations are quasi-static; others have an uncalibrated time scale; the rest have limited dynamic range. Multiple models and other (non- SM) approaches such as historical analogies and precursor analyses can prove useful where applicable. 4) Digraphs Are at Least as Important as Matrix Repre- sentations When Direct Connections Are Subjectively Esti- ‘mated: Faced with a matrix format, the user is prone to fill in most of the possible entries. The connections, how- ever, are often indirect, a fact more easily recognized in a digraph. ‘The widespread use of matrices in the development of structural models has led to many modeling errors. Unless the number of variables and links is very large, we recom- mend that digraphs (rather than matrices) be used in the early stages. They can always be converted to matrices for entry into computer programs if necessary. We recommend that the person considering the use of SM first attempt to develop # digraph of at least a subset of the elements of concern. The behavior of the system may be inferred by inspecting feedback loops and the connections involving each element. If this attempt fails, or if the digraph is difficult to interpret, then it may be useful to engage a consultant skilled in SM. If the initial experience proves fruitful, then it is possi- ble to determine whether SM consultants should be drawn in to help build structural models with a large number of elements (e'g., Spin or ISM). Once consultants have pro- duced worthwhile results, then consideration might be given to developing in-house capability for carrying out subsequent analysis. Learning-by-doing is a valuable way to proceed in developing application skills with SM tools. 5) Using ISM as a Precursor to KSIM, SPIN, or Any Method Using a Causal Digraph as Input May Lead 10 Errors: ISM not only assumes transitivity but by its na- ture does not address many potential connections. When it finds a cycle, it chooses a proxy element and all connec- tions to any element in the loop are shown as connections to the proxy clement. Connections must be added to other elements (and possibly subtracted from the proxy ele~id IEEE TRANSACTIONS ON SYSTIMS, MAN, AND CYBERNETICS, VOL. 54C-10, NO. 12, DECESBER 1980 ment) to correct this problem, a task which may be more difficult than developing a causal digraph from the outset. Our experience indicates that ISM users evince a cer- tain conservatism. Whereas “yes” denotes a clear cut relation between elements, “no” often reflects only an absence of such clarity of relation. The resulting digraph may prove difficult to convert to a causal digraph. In any case, we know of no reliable shortcuts to the development of a causal digraph, nor do we know any reliable means of converting a digraph to the appropriate causal digraph for the same situation. 6) Selecting the Appropriate SM Tool From Among the Seven: ISM and Spin-are most useful in dealing with a large number of elements. Both are easy to use. KSIM is, ‘most useful when there is a small number of highly aggregated variables. It is relatively easy to use. QSIM and ELECTRE are also useful and provide credible re- sults, but their spectrum of versatility is narrower. Impact is easy to use and may produce some insights, but the results are less credible. Fig, 5 is a heuristic guide for selecting from among the selected SM tools. The reader is reminded, however, that SM encompasses many other tools (cf. (63] and Table IID). ‘Their exclusion here is not meant to imply they are useless. Tt has been postulated that, in modeling, the most successful tools tend to be the most problem-specific. Iti, in fact, their suceess in that specific context which often, Jeads to their generalization, then their inclusion in a “tool kit.” subsequent misuse, and finally disrepute. The user must be cognizant of the dangers at all times and con- sciously suppress a rigid, uncritical attitude vis-a-vis his or her “favorite” tools. B. Needed Work for Improved Use of SM The profusion of tools of limited conceptual range suggests that the SM effort to date has been estensive rather than intensive. A few researchers, such as Harary et al. (38}; Warfield [100}, (101), [102]; Roberts [81]; and Cearlock [13] have pushed to greater depths—but more needs to be done. 1) Better Documentation: A detailed workbook or man- ‘ual is needed which describes step-by-step how to build a structural model and how to use specific SM methods. It must also provide examples. The workbook should be oriented toward manual analysis and should not push the user to rely on computers in every application. 2) A. Superstructure?: The model builders have pre- sented us with a plethora of tools which confound the nonexpert! In their zeal for their respective methods, model builders often broaden the original tool applica- tions to achieve wide appeal as generalized problem- solving procedures. They exacerbate the user's difficulties by continuing model “improvements.” Often only the model developer is able to implenient and apply the latest version. This pattern is reminiscent of the “planned ob- solescence™ principle pioneered in the auto industry. Part ‘of the problem has been the lack of coordination among structural modelers: ‘The history of use of structural models is on¢ of lack of communication, lack of cumulation of knowledge, and inadequate sharing of experience. This is particularly well illustrated by the history of development of KSIM, on the ‘one hand, and QSIM on the other. As originally devel- ‘oped, each had shortcomings. Each was further refined to overcome these shoricomings. As a result, we have a proliferation of models, each with its own users and its own peculiarities, and no general knowledge in the com- ‘munity of what a good model ought to look like {65} One way to ameliorate this dilemma would be to design and implement a superstructure that would provide com- ‘mon input, storage, editing, and output functions to all the computer-based tool algorithms. This superstructure would help reduce redundant capabilities of the present tools, would help to standardize procedures, and would help to improve the man-machine interface. (Spin, which is far superior to any of the other tools vis-a-vis this superstructure concept, could serve as a model for this design.) In addition to the data management duties, the execu- tive routine of the superstructure could also serve as an interactive conversational “front-end” implementation of Fig. 3 (the “Map"), guiding the user through the assump- tions he is willing to make about the system he is model- ing and the level of information that he is ready to supply to the model. This conversation with the front-end soft- wware would help select a suitable algorithm at that time. ‘Another feature that could be built into this integrated structural modeling software package is a random number generator along with an automated rerun of the model With new random numbers. If feasible, this would provide the user with a chance to run the structural models in a Monte Carlo mode, yielding a probability distribution of outputs which at times would be more appropriate to SM (given the uncertainty of judgments that go into a struc- tural model) than the single output now yielded by all the tools. Algorithms could be added to the library. System dy- namics could be included, serving as an end-point for those desiring to take their modeling efforts all the way through to dynamic modeling. 3) Improvement of Existing Tools: Research would be desirable to improve the human engincering of ISM and Spin. Better group input and graphical output techniques are needed. While relevance trees have already been applied suc- cessfully in broad problem contexts (as stockpiling and sgeothermal-energy technology-assessment projects), varia- tions such as problem-solution trees, mission flowcharts, and social networks [63] have not yet been exploited. Their potential makes it desicable to develop them further. Tntegration of Spin and QSIM to develop their inter- activeness is desirable. For example, a computer program ‘could query the user to establish additional information: Which variables are ausiliary? Are sophisticated couplings between variables needed? Credible time trajectories can-EDANIS: STRUCTURAL MODELING Fig. 5. Guide for choosing among recommended SM tools. bbe produced in this way, in addition to the initial struc- tural analysis. 4) Basic Research: : Causal graphs: Methods are needed for developing causal graphs to represent specific systems or problems. One approach is to develop a transistive digraph using ISM and then to convert it to a causal digraph. Research may determine whether a conversion process can be iden- tified and automated. Once a causal digraph is available it may be possible, judging from current research, to use algorithmic methods for creating a dynamic model (11]. Stability and resilience: Our understanding of the ef- fect of new technology is minimal insofar as it affects the domains of stability in societal systems. Measures of resli- ence need to be developed, and the safe-fail concept (40], of great significance in social-system analysis, must be Placed on a firm footing. 5) Man-Machine Interaction: One of the most promis {ng new possibilities in connection with SM is a procedure to do the structuring process in an interactive mode using 2 CRT (g,, Langhorst [55). Using a lightpen a sketch of the digraph of the system is entered on the terminal. The ‘computer requests necessary parameters and initial condi- tions, performs the appropriate mathematical operations and displays the results on the terminal (Hudetz (43). We are reminded that +++ the computer can give the individual a new ability to play with very complex systems, to get the ‘feel’ or ‘experience which he relies upon to comprehend complex systems. We refer to an informal, intimate dialog, the concretization of Gedankenexperimente. The key is the vse of multiple sensory modalities in interacting with the ‘computer (tanual-visual, audio-visual) (61, p. 57}. Substitution of numerical output by words or symbols should be considered to reduce the temptation of misuse of hard-copy output. Examples: words—“oscillate,” “in- crease slowly,” symbols—arrows of varying thickness andoe af scaled size circles (eg, size may indicate number of sy on Such graphic output would minimize the misinterpreta- tion of pulse plots of Spin as a time series or the time series plot of KSIM as specific quantitative forecasts For the process of learning, the impermanence of a cathode ray tube image is far more appropriate than hard-copy output, just as chalk and blackboard are useful in the classroom. The GRIPS SM tool output is, in fact, CRT graphics-oriented and has a unique advantage in this regard (see [43)). 6) New Structures: Perhaps the sharpest blow which can be leveled at SM is its preoccupation with a minute class of representations— trees, digraphs, networks, matrices. They suggest a woeful lack of imagination, a “flatland” inability to conceive of more than two- dimensional space. Why should such complex systems be reduced so uniformly to digraphs? Are these the only simple structures we can envision? Complex systems prac tice self-organization and create new structures; a spiral ‘or a helix with few elements might tell us more than a hierarchy with many (62, p. 258]? Prigogine’s “order through fluctuation,"* and Holling’s “safe-fail” stability domains may be leading us toward structural modeling more appropriate to complex human systems. We suspect that addition of societal and behavioral perspectives to the problems being tackled will require modeling tools of a different genre. This represents a challenge to which we must rise if we of the “systems” community are to provide real assistance in solving the complex problems facing our society. Actually the emphasis on hierarchical structures is a striking reflec: tion of a thought pattern characteristic of Western cultre. The Limited range of structures 40 far considered underscores the dominance of the principle of homogencity. More attention to other eutural patterns, based, for example, on heterogeneny and mutual causality, might gener- stg quite difereat structural models. eg. see work of Maruyama (67) ‘Fluctuations lead to uapredictable changer (analogous to biologial ‘mutations) which create a new, sable, and very dillerent—pombly ‘more advanced—system structure [95], With time, fluctuations may begin agein, become mote pronounced, and repeat the eyele. Such aa evolutionary, seliorpanizing process should be of utmost significance, Dt cannot be analyzed withthe current sate of the art in SM. [EEE TRANSACTIONS ON SYSTEXS, MAN, AND CYBERNETICS, VOL, S4C-10, NO. 12, DECHAGER 1980 1 ©) A Final Thought:. The structural modeling process serves to expand our thinking about complex system be- havior. Such a process is a means to reach and perhaps augment our intuition; there remains, however, a vital distinction between the latter and hard copy “results” which might be generated. ‘Thinking is enhanced by the organizing, analyzing, and ‘communicating functions of SM. Even a crude model may prove useful: a quick and inexpensive means of probing can become very productive and cost-effective in the right hands. ACKNOWLEDGMENT The author wishes to acknowledge the considerable contributions of the other members of the project team, in particular W. Wakeland, S. Rogers, M. Williams, and the team leader, H. A. Linstone (see (63)). A majority of this paper was prepared while the author was on sabbatical leave at Stanford University, Department of Enginecring- Economic Systems, and he wishes in particular to ac- knowledge the support of that institution for the prepara- tion of the extensive graphics in this paper. REFERENCES U1) C. Alexander, Notes on the Synthesis of Form. Cambridge, MA: Harvard, 1964 [B}_B. Allen, "GENSIM—a generalized digraph simulator” Rutgers, Univ., New Brunswick, Ni, 1975 B) 8. Alte, “The comptauonal mathematics of time dependent crossimpact modeling,” Center for Futures Research, Univ. Southern Califor, Los Angeles, CA, Tech. Rep. M26, 1976. 14) RM. Axelrod, Ed, The Structure of Decision, Princeton, NJ: Princeton Univ, 1976. (5) M. M. Baléwis. Bd, Ponraits of Complexity, Battle Memorial Tastiute, Columbus, OH, Battelle Moaograph No. 9, 1975 16) M. Bloom, “Deterministic trend crosvimpact forecasting” Teche tol, Forecasting Social Change, vol. 8, pp. 38-74, 1975. 2 program for evaluation of the effects of interacting random eveats and regular phenomena on a fore ‘cea development,” Reo. Teck, Thomson CSF, vol. 6, pp. 661-681, 1974 (a Freach). 8) G. Brocksta, "Constraint analysis and structure identification” Annals of Systems Res. v0.5, pp. 61-80, 1976 (9) 5. R Burns and O. Ulgen. "Adjacency matrices and system dynamics” presented at Pitsburgh Conf, Modeling and Sumula- ‘on, Apri, 1977 (10) J. R. Burs and W. M. Marcy, “Causality: Is characteristics in ‘system dynamics and KSIM models-of socioeconomic systems.” Dep. of Systems, Texas Tech. Univ. Lubbock. TX (to be pub- lished in Technol. Forecasting and Sve. Change) [11] JR, Burns and O. Ugen, "A sector Gubsystem) approach to the Tormulation of systems dyzamics models,” 1977 (unpublished. 2} RUE. Cavalo, "Recosstrictabilty analysis of multidimensional ‘elation: a theoretial basis for computer-aided determination of {eceplable systems models,” In. J. General Systems, vol. 5, 20.3, pp. 2-171, 1979, . B. Cearlock, “Common properties and limitations of some structural modeling techniques” PLD. disserauon, Univ. ‘Washington, Seatde, WA, 1977 “A sirctural modeling technique using information 003) 4) measures” in Proc. 24th Anns. North Amercon Meeting of Soiey 115) °K. Craver, Croselnpact Anahi. Technology Astesrment in Procice. St Louis, MI: Monsanto, 1972. 6] N. Dalkey, “Aa elemenuary crossimpact model,” in The Delph Method: Techniques ond Applicauons, H. A. Linstone and M. ‘Tarot Eds, Reading, MA: Addison-Wesley, 1975a ‘Lewpanus: STRUCTURAL MODELINO 07 us 119) 0) Ry 3 co 4 es 6 27 a) 91 0) bu 2) 64 1s) ba om 1a) (4 ta) 14) (1 a P. Delp, A. Thesea, J. Motvalla, and N, Seshadsi, “Systems tools for project planning.” Program of Advanced Studies ia Inst: toa Building and Technology Asistance Methodology, Indiana Univ. Bloomington, IN, 1977 RLA Dudek er af, “Technology Asvessment: Human Rebabilta- ‘lon Techniques,” Texas Teck. Uaiv., vol. VI (undated). J.C. Duperin and M, Godet. “SMIC 74—a method for coastruct- ing and ranking scenarios," Furwrer, pp. 302-312, Avg. 19758, ——, “Conjectural systems analysis: problems aad approach,” ISTA J. (at Soe. Techaol Assessment), Dee. 1975b. S. Enier, “Some progress aad problems in technology assestment TA statement before the Congressional Board of the Office of ‘Technology Assessment.” Cente for Futures Research, Graduate Schoo of Business, Univ. Souther Califor, Lo Anges, CA, 1976, W. R Ferrell and R. H. Miligan, “An experimental tet of KSIM: a model for aiding intuitive forecasting.” preseoted at 1975 IEEE tat. Conf. Cybern. and Soc, San Francisco, CA. 1975. Bro R. Fit, “loterpretve sirvctural modeling as techaology for socal tearing.” in Portraits of Complexity, M. M. Ualdwan Ed, Batcle Memorial Institute, Columbus OH, Bauelle Monograph No, 9, 1975, Bro. R. Fit and J. Troha, “Interpretive structural modeling and urban planaing.” Proc. IEEE Int. Conf. Cybern. and Soe, O%. 1977 Bro R Fitz and Horbach, “A participative methodology for designing dynamic models through sinuctural models” Prov IEEESMCS Int, Conf. Cybern. and Sec. Oct. 1977 43. W, Forester, Principles of Systems. Cambridge, MA: Wright: ‘Alea, 196 —, Werld Dynamics. Cassbridge, MA: WeightAllen, 1971, J. Fried and P. Molaar, “A general model for culture and technology,” Technol. Forecasting Soe. Change, va. 8 pp. 175 u Fogmaaa ef al, “Represcatation of concept relations wsing the TOSAR system.” J-'Amer. Soc. Information Sci, vol. 25. pp 287-301, 1974, L. Gerardi, “To forecast decision-making with systems analysis, ‘temic analysis to foresce alternative fstares,” Thomson-CSP, Paris, paper LG/JP-DTG-523/74, 1974 —, "Meses of problems: conflicts in mesh,” Thomson-CSF, Paris, paper LG/DA-DTG:532/76, 19763. —, “To creatively invent futures in agreement with technology {Assessment through mompbological analyns put under structural ‘model constraint” pressnted at Second ISTA Congres, Univ. MI, Oct 24-28, 19760. i, Structural Modeling ot an Aid to Assess the Releoance of Conflicting Policies and’ Sirategies of Social Actors, Pass: ‘Thomsoa-CSF, 1977. T. 4. Gordon ‘and H. Hayward, “Inial experiments with the roubimpact matrix method of forecasting.” Purves, vol. 120. 2 Dee, 1968. RJ. Gordon and J. Stover, “Using perceptions and data about the future to improve the simulation of complex systems,” Tech ing and Soe. Change, vl. 4, pp. 399-408, 1579. KC Hammoud eral, "Social judgment theory,” in Human Jud iment and Decision Processes, M. Kaplan and S, Schwartz Eds New York: Academic, 1975, ch. 10, F. Harary, R. Z Norman, and D. Cartwright, Structural Models: ‘An Introduction 10 the Theory of Diected Graphs. New York Wey, 1965. K. Haruna and N. Komoda, “An algorithm foc structural seasi- LWvity analysis in structural analysis” Proc. [EEE Int. Conf ‘Spuiems, Mam, and Cybern. (Japan) 1978, 07 Helmer, "Problems in futures research” Funes, pp. 17-31, Feb, 1977. C.S, Holling, “The curious behavior of complex systems: lessons from ecology." in Futwres Rezearch: New Direchons. WA; Linstone and W. H.C. Simmonds Els, Reading, MA. Addison Wesley. 1971. pp. 114-129. C.S. Holling and M. A. Goldberg, Resource science workshop.” Resource Science Ceatte, Univ, Bntsh Columbia, unpublished, W. Hudet “A graphical nteactve sysiem simuliion progam fot mini-computers,” ia Concepts and Tools of Computer Assisted Policy Analysis, H. Bosel Ed.” Basel: Bukhauser, 1917. J. Kane, “A primer for a new crose-impact langusge—KSIM." a 39 Technol. Forecasting and Soc. Change, vol. 4, pp. 129-142, 1972 J. Kane, W. Thompson, and 1. B. Vertissky, “Health care de- livery: « policy simulator” Socio-Economic Plamsing Sci. vol. 6, pp. 283-153, 1972 J: Kane and LB. Vertinsky, “The arithmetic and geometry of the tuture” Technol. Forecating and Soe. Change, vol, p19, 1975 K. Kawamura and A. N. Christakis, “The roie of structural modeling in technology assessemeat” presented atthe 2nd ISTA Congress, Univ. Michigan, Oct. 24-28, 1976 Kehoe and. Clif, “INTERORD— A. computer interactive FORTRAN IV program for developing simple order," Face: onal and Peychol. Measurement, vol 33, pp. 615-678, 1975 G. Kir, “Ideatfcation of generative sructuses in empirical data” Int J. General Sy, v0. 3,20. 2, pp. 89-108, 1976. G. Klir and H. 4.5. Uyteshove, “Oa the problem of computes Aided observations and resulting guidelines,” Int. J. Man-Machine Saies, vl. 9, 1977 K. Krippendort, “A spectral analysis of relation, funher devel. ‘opments. in Prac. Fourth European Meeting on Cybern. and Sys. Research, F Pichler, Ed, Austria, March 1978. P.G. Kise, "KSIM techniques for evaluating interactions among, variables” Stanford Research Institute, Mealo Park, CA, Tech. Note OED-16, Juse 1973 —, “Cross impact simulation in water resources planaing,” Stanford Research Insutue, Mealo Park, CA, 1974 P.G. Kruzie and 8, Sandys, KSIM Computer Manual, Tech. Now CRESS TNI6, Stanford Research Institue, Washington, DC, or FE Langhorst, “Computer graphics aided interpretive structural ‘modeling? a oo or conceptualizing complex design problems, Ph.D. isseratioa, Pardue Univ, Aug. 1977 G.G. Lendars, "Oa the buman aspects of structural modeling.” Technol. Forecasting and Soc. Change, vol. 18, pp. 329-351, 199, —""Some coasiderations ia the we of ISM," 1979, submited {or publication. W. ‘Leoatiel, Structure of the American Economy, 1919-1939. New York: Oxford Univ, 1951, 2nd ed. —, Input-Output Economics. New York: Oxford Univ. 1966, Levins, "Evolution in communities near equiibsum.” a Ecol. omy and Evoluion of Communities, Cody and Diamond Eds Cambridge: Belenap, 1975, pp. 16-50. “Four Amencan futures: reflections on the role Technol. Forecating and Soe. Change, vol. 4, PP 45) 46) “ 19) isn (32) 13) 14) 155) 156) 1 (58) (39) (oo (oy (a (o HA Linstone and W. H. C. Simmonds, Eds. Futures Research ‘New Directions, Reading. MA: Addion-Wesley 1977 HUA. Listone et a, “The use of structural modeling for technol- ofy assessment” (2 volumes), Futures Research Iasutute, Por aad State Univ, Pordand. OR, Publication 78-1, 1978 [604] —, “The use of structural modeling for technology astess ‘meat™ (summary article version of [6)), Technol. Forecasting and ‘Soe. Change, vol. I pp. 291-327, 1979. C. Marchet, "Primary energy substitution models: 09 the inter- action between energy and society,” Technol, Forecasting ond Soe. ‘Change, vol. 10, pp. 343-386, 1977, J.P. Marino, presented at EIES Conference Comment entered ‘on Electronic fnformation Exchange System, Coal. 318, Com ‘meat 14, Dec. 31, 1978, W. Marcy, “OSIM-a modified cross-impact language.” Dep. Industrial Engineering, Texas Tech Univ, Lubbock, TX, Tech. Rep. QT-110-16, 1976, M. Maruyama, “The second cybernetics: deviation-amplifying mutual causal processes” ia Modern Sputems Research for the ‘Behacioral Scemist, Buckley, W. Ed.” Chicago: Aldine, 1968, J.-M, MeLean, “Getting the problem right—a roe for strctaral modeling” in Furres Research: New Directions, H. A. Linstone od W. H. C, Simmonds Eas. Reading, MA, Addizoa-Wesey, 1977, pp. 144-157 J. M. McLean and P. Shepherd, “The importance of model Sovctue,” Furwes, Feb. 1976. JM. McLean, “The United Kingdom private car model—an ‘sample of structural modeliag.” 19773, uapublshe. ILM. McLesa and R Cumow, “Techniques for analysis of system structure,” Sciece Policy Research Unit, Univ. Sussex. Falmer, Brighton, Sussex, United Kingdom, SPRU Occasional Paper Series No. 1, 19762 J.-M, MeLean, “Progress ia structural modeling —a biased re- view.” Science Policy Research Unit, Univ, Susex, 1976 ( (651 (66 to (6s (or (704 mm (my0 (73} RB. Mitchel and J. Tydeman, “A note on SMIC 74," Funwes, PP. 64-66, Feb. 1976, 174) A. Mitchel al, “Handbook of forecasting techniques.” Stun- ford Research Insitute, Mealo Park, CA, Policy Research Report 4a, 197s. RH. H. Moll and C. M. Woodside, “Augmentation of coss- {impact analysis by iterative systems analysis tools.” Dep. Sys tems Engineering, Carleton Univ, Ottawa, Canada. Rep. No. SE-61, 1916. 03) (99) oy SLL LLL seas [EEE TRANSACTIONS ON SYSTEXS, MAN, AND CYBERNETICS, VOL. SMC-10, NO, 12, BECHLER 1980 (Cybernetic Systems Program, San Jose State Univ. San Jose, CA, 17 P. Sbepberd and M. McLean, “The SPIN System” Computer Software tipes SPIN-DOC, Science Policy Research Unit, Univ. ‘Sussex, Falmer. Brisbton, Sussex, United Kingdom, De. 7, 1976 LC Silvera, “LOGOS—a system language for flowchart model. ing.” Educational Teckel, vol. 9, pp. 18-23, 1969. BE. Suta, “KSIM theoreieal formulation: «parametric analysis” Stanford Research Institute, Menlo Park, CA. Tech. Rep. TN- (16) N. O'Leary, Forecasting with Dynamic Systems. New York O£D-25, 1974. Learning Resources in International Studien, 1975, [92] P. Tawfik, “An improved approuch for “optimization” of multiple (77) C.L, Owen, “An algorithm for the decomposition of nondirected policy objectives." Pa. disertation, Porland State Univ, Port. ‘saphs.” in Emerging Methods in Enoironmental Design end Plan land, OR, 1977, ting, G. . Moore, Ed 1970. [93] M. Turoff, “An alternative approach to eros-impact analysis” in (78) J.'L, Piala, “Graph struetures.” J, Associaton for Computing The Delphi Method: Techigues and Applications, H. A. Linsione ‘Machinery, vl. 18, pp. 411-42, 1972. and M. Turofl, Eds. Readiog, MA: Addison. Wesley, 1975. (09) J. Randers, "Methodological aspects of social stems simulation,” [94] §. Umpleby, “The Delphi exploration” a computer based sytem ont. D. Meadows Ed, Conf. at Sukhumi, Abkhasia, USSR, {or obtaining subjective judgements on altervative futures," Univ. Jan, 25, 1974. linois, Urbana, IL, Tech. Rep. Fl So} F. E. Roberts, “Weighted digraph models for energy use and air [95] H. vou Fons, “The curious bebavior of complex systems: pollution in transportation systems,” RAND Corp, Rep. R578 lessons from biology." in Futures Research: New Directions, H. A. NSF, 1974 Linstone and W. H.C. Simmonds Eds. Reading, MA: Addison. (81) —! Discrete Mathematical Modes, Englewood Cills, NI Wesley, 1977. Prentice-Hall, 1976. (96) W. Wakeland, “QSIM2: & low budget heuristic approach to 82] F. E. Roberts and T. Brown, “Signed digraphs and the energy ‘modeling and forecasting,” Technol. Forecasting ond Soe. Change, tsi” Amer. Math. Monty, 8216, pp. 571-594, 1975, vol. 9, pp. 213-229, 19760. 5} B. Roy, La Methode ELECTRE If, Une Methode de Classement (97) ——, “ASSESS: an application of structural modeling to technol en Presence de Criteres Maliples, Group METRA, N. T. No. 142 oy Asestmeat.” Systems Science Ph.D. Program, Portland State Qs. 1971 Univ, Pordand, OR, Paper 76-1, 19766. (4) B. Roy and P. Benet, “La Methode ELECTRE Ml, Une Applice- [98] W. Wakeland and E. Solley, QS/M2 User's Memul, The Lile ‘ion au Media-Planning.” in Proc. Sith IFORS Int. Conf. Opera- Support Systems Group, Lid, Portland, OR, 1977, ional Research, Avg, 21-25, 1972, Dubin, Ieland, pp. 291-302, {98} R. Waller, “An application of interpretive rvetral modeling to 1973, Priority acting in urban systems management” ia Portals of [85]. Sahat, “Conception of futures in a systems framework,” ia Conplesin, M. M. Baldsan Ea, Battele Memorial Tasttute, Futures Research: New Directions, H. A. Linstone tad W. H.C. Columbus, OH, Battelle Monograph No. 1975 Simmonds Eds. Reading. MA: Addison-Wesley, 1977. (Hoo) 5. N. Warfield, Am assault on Complexiy, Battele Memorial (86) —. “A theory of measuement of technology.” Int. J. of Sy. Institute, Columbus, OH, Battelle Monograph No. 3. 1973. tems ‘Sct, 1977 Ol] —, Sinsturing Complex Systems, Batele Metoral tastnte, (87 —, “Towards a theory of systems” in General Systems Columbus, OH, Battele Monograph No. 4 1974 Yearbook. 1977 [N02] Societal Sytems: Policy, Planning, and Canplesty. New 185] If had @ Hammerand Other Tools—A Systems Tools Handbook, George G. Leodarts (S'SS-MS8-SM74) was ‘born in Helper, UT in 1995. He received the BS, MS. and PhD. degrees in clectncal em: incering from the University of California, Berkeley, in 1957, 1958, and 1961, respectively. From 1961 to 1969 he was a Senior Research Engineer at the Geneva! Motors Defense Re- search Laboratories, Goleta, CA. Research Be was interested i inteligence and pattern recognition. He iatro- ‘uced the notion of and did exteorve research while thre iacluded artificial ‘on optical power spectrum sampling for pattero analysis. His praciple paper in this area was selected by IEEE for inclusion in the IEEE ‘Selected Reprint Series on Computer Methods in Image Analysts, 1a 1969 ‘be weat into academia, going fist to the Oregon Graduate Center for study and research where he was Chairman of the Faculty and sared developing a Systems Science Department He moved to Pordend State Univenity in 1971 to join the then new Systems Science Ph.D. Program as 2 Professor of Systems Science. He has been instrumental im the development of that fourdepertment, interdisciplinary program, His ‘esearch interests are now concerned with development of methodologies {or assisting teams of people to carry out systemrelated tasks (Geren, ‘engineering, etc). Strvctural modeling is one of various appreaches in (his context Dr. Lendars is a member of the Society for Geveral Systems Re- search, Pattern Recognition Society, Assocation for Traaspersonal Psy: chology and Sigma XI. He ison the editorial board of the International Journal of General Systems. He has served the IEEE Systems. Man, tnd ‘Cybernetics Society in several capacities: Chairman of the Education Committee, member of the Administrative Committe, and Vice res deat for Publication. He has served in government in various advisory ‘capacities: Oregon Governor's Technical Advisory Committee (1970 1972), Oregon State Senate Task Force on Eeonomic Developmeat (1972-1973), Multoomah County Economic Development Advisory ‘Commission (1977-1979) and the Economic and Fiaancial Advisory Pane, Portland Energy Commission (1980)