Connecting expert system features to a multiple criteria programming based decision support system

175

Computer Science in Economics and Management 1 (1988) 175-188. © 1988 by Kluwer Academic Publishers.

Connecting Expert System Features to Multiple Criteria Programming Based Decision Support System

a

R A L F O S T E R M A R K and H A N N U S A L M E L A Abo Academy, Henriksgaten 7, 20500 Abo 50, Denmark

(Received: 15 March 1988; revised: 15 June 1988) Abstract. We concentrate on the issue of utilizing artificial intelligence based computer systems as an aid to formulating and formalizing corporate planning problems. Our proposal involves cooperation between the analyst and expert system in order to find a suitable structure for the problem to be solved. The A1 system is designed using logic programming. The study is an outgrowth of the experience gained in developing a strategic planning system with menu-driven simulation (heuristic, Monte Carlo), multiple criteria optimization and database query facilities for the concern management of Kansallis-Osake-Pankki (Finland).

Key words. Decision space, managerial support, logic programming.

1. The ,Corporate Planning System The idea of using artificial intelligence to support corporate planning was originally born as a by-product in the development process of a traditional and formal corporate model. In order to give a more comprehensive view of the role of AI systems in corporate planning, the key functions of the corporate model will be presented. (For a more detailed description of modelled system, see [16].) The corporate planning system is based on a set of integrated models, the output of which consists of balance sheets, income statements and some special reports. The reports generated using the system include actual numbers for a couple of history years and forecasted numbers for the planning period. The general structure of the report module is presented in Table I. The system allows the user to generate forecasts by two alternative methods, simulation and optimization. While developing the system, these two methods were not seen as competitive but supportive ones, both having their justification in solving certain types of problems. In simulation, the first step is to generate a base forecast. The initial forecast can be tested with temporal changes by using various facilities such as W H A T IF, IMPACT and G O A L SEEKING. If some of the assumptions can more conveniently be given using ranges or distributions, there is a possibility to make stochastic computations with Monto Carlo simulation. Forecasts made for the single corporations of the bank concern can be consolidated into one forecast for the whole concern. The consolidation is performed according to Finnish legislation.

176

RALF OSTERMARK AND HANNU SALMELA Table I.

The layout of system generated reports.

History years . . . 1985

Forecast period 1986

1987

1988

Income statement (Actual)

Income statement (Forecast)

Balance sheet (Actual)

Balance sheet (Forecast)

Special reports (Actual)

Special reports (Forecast)

1989

1990...

A critical aspect in simulation is the number of assumptions and growth factors that must be determined before the forecast can be computed. Simulation is especially suitable when comparing well-defined courses of action, since most of the assumptions can be given meaningfully and the resulting forecasts computed in no time. In unstructured planning situations, however, such alternatives do not exist, but one should be able to make a realistic plan by considering both the relevant constraints and the conflicting objectives in problem solving. In these situations simulation may turn out to be quite a cumbersome method for making a forecast, as several trials must usually be done before all relevant factors have been taken into account. In order to strengthen the system capabilities in these cases, the possibility to optimization was included. In optimization, the search process for an, acceptable solution is done by computer. The user must still state the constraints, in order to obtain solutions that can be accomplished in practice. Since there will mostly be more than one possible solution, the user must also articulate a preference order over the objective set. Basing its search process on this information, the search algorithm seeks for the best solution by methods of multiple objective linear programming. If the problem definition has been given properly, the solution selected is simultaneously • • • •

realistic with respect to the market situation, legal, concordant with managerial policies, efficient with respect to the objective set.

An advantage of using optimization instead of simulation is that various forecasts can be made in a short time. After the problem definition has been made, new and totally different forecasts can be generated only by changing the preference ranking of the objectives. But optimization naturally has its drawbacks as well. Initially, it may be difficult to notice all the relevant constraints in certain problems. The preference of the objectives requires several trials, before the trade-off possibilities between objectives are adequately explored. Thus, the preparation of the first single forecast usually takes

CONNECTING EXPERT SYSTEM FEATURES

177

longer than, for instance, in simulation. However, subsequent forecasts can be generated considerably faster and, furthermore, the resulting forecast is more closely argued because of the detailed process required to produce it.

2. Need for Intelligence in Systems Use The usefulness of traditional operations research tools for an ordinary user - even if menu driven - can be questioned. The concepts used in optimization are mostly unfamiliar to potential users. The users are normally the best experts on their specific problems, but only rarely have they ability to express their problems in the form required by traditional OR models. Even if general usability problems can be seen as tasks for artificial intelligence [1, 10], in this paper we will concentrate on problems specific to optimization. It seems likely that before optimization can be used as a decision aid in unstructured problems, we need to be able to assist decision-makers to define their problems in a way which supports subsequent formalization efforts. The identification of the proper variables, relations, actions, constraints and objectives from the problem description requires linguistic interpretation. A possibility might be the use of artificial intelligence in supporting the decisionmaker in the early stages of problem structuring. It has been stated that both structured [12] as well as unstructured [5] decision processes can be more conviently described to computer using rule based programming. However, even with rule based programming the task is nothing but easy. In the field of management, much of the most important knowledge is quite resistant to standardization and pre-specification [14] and knowledge can be fully captured only in natural language [9]. Before proceeding to the actual use of artificial intelligence, we shall shortly describe the need for intelligence from a practial point of view. In this section, we present the way how problems should be introduced to the corporate planning system previously described. Planning by optimization can be divided into four phases: (1)

(2) (3) (4)

Definition of the scope of the planning problem by naming the decision variables, for which optimal values are to be found. Analogously in order to gain a refinement of the simulated forecast, it may be utilized as a basis for optimization: a subset of the input variables may be declared as decision variables in the optimization stage. Definition of the solutions that are accomplishable in practice by stating an adequate set of constraints. Selection of the best alternative according to the preference order imposed on the objectives. If the preference is known, it is possible to select a compromise. Exploration of the solution and return to some of the previous phases if the solution is not satisfactory.

178

RALF OSTERMARKAND HANNU SALMELA

Research concentrating on multiple criteria decision methods has traditionally been concerned with preference ranking (phases three and four). From our point of view, however, the critical phases are the preceding two, where the problem is actually defined. Thus, in this paper the main emphasis will be on these phases. Strategic planning is an unstructured task, hence soft assumptions and judgement are required in the problem identification/definitionprocess. The user should have the possibility of altering the decision variables and constraints selected during the forecasting process. In the planning system described, the problem definition is written to a specific directive-set file. Both decisions and constraints are given as simple clauses in the following form (required by IFPS, the financial planning language used): DECISIONS PRODUCTION 1 GROWTH (3-6) PRODUCTION 2 GROWTH (3-6) CONSTRAINTS INVESTMENT ON PROD1 CAPACITY (3-6) .LE. 1000 INVESTMENT ON PROD2 CAPACITY (3-6) .LE. 300 MAX GROWTH IN SALES1 (3-6) .LE. 2000 MAX GROWTH IN SALES2 (3-6) .LE. 800 The figures in parenthesis denote the planning periods. Thus, in the planning system, the problem definition is made simply by editing the directive set file. Technically, this is not a difficult task, provided that the user has some experience in using a file editor. The difficulty lies in the fact that alterations of the problem definition require expertise on the logic of the actual problem that should be solved as well as expertise on the formulation of problems. The optimization algorithm used in the system imposes two mathematical requirements on problem formulation: • there must be an adequate set of constraints on the decision variables such that the solution space remains feasible; • the problem formulation must be linearized, implying that the decisions and constraints must be carefully designed in order to avoid nonlinearities and/or infeasibility. Efficient tackling of the feasibility issue requires detailed knowledge on the actual model to be optimized and ability to comprehend the principles of the optimization algorithm. This knowledge is normally expected from an analyst, not from an ordinary decision-maker. Neither has an ordinary user expertise on linearizing a problem formulation. Even if the rules for linear representation are not impossible to learn, the use of the trial and error method in altering the problem formulation may prove troublesome.

CONNECTING EXPERT SYSTEMFEATURES

179

Still, the transformation of the decision-maker's experience on the actual problem into a problem formulation is prerequisite to optimization. In principle, the following three ways were considered for performing the transformation: (1) Preparation of alternative pre-defined ways to change the original problem formulation made by the analyst. In practice, this would require several predefined concepts from which the user could select a suitable one. (2) Cooperation between decision-maker and analyst, in order to identify the problem. The cooperation may be organized in different ways, but it should take place both a priori and during the actual forecasting process. (3) The use of artificial intelligence to support the decision-maker in problem formulation, that is, quantifcation of the relevant criterion and decision spaces. In the second alternative, the problem formulation process is guided by a human expert. In the third alternative, the human expert is replaced by an expert system. The possibility of replacing an analyst with an expert system is well reported by Hendry [3]. He states that there are several weaknesses in expert systems, thus they are only able to support other disciplines, a conclusion which is very similar to our view. From the viewpoint of DSS- and MCDM-oriented planning systems, the possibility of using artificial intelligence in formulating the criterion and decision spaces is an interesting alternative.

3. Formulation of Criterion and Decision Spaces A fundamental assumption of our approach is that the performance of financial management can and should be improved in a computer-aided multiple-criteria decision making (MCDM) framework. A graphical decision support system (DSS), based on the idea of a multiple criteria conflict zone encompassing all efficient solutions, was originally presented in 0stermark and Kasanen [17]. The DSS- and MCDM-dimensions of the proposal were investigated further in [6, 7, 18-22]. The core idea of the conflict zone approach is to successively narrow the space of nondominated solutions, through articulated aspiration levels for one or more criteria in the criteria set. In order to overcome the problem with underestimated lower bounds [13], the initial conflict zone is constructed with logical minima imposed on the decision space. Such logical bounds are obtained from, say, the lower median values for the critical success factors of the branch to which the specific firm belongs. The conflict zone framework is specifically designed for interactive use, supporting rapid convergence to managerially meaningful aspiration profiles within the initial zone (desired directions in the criterion space). Efficient solutions that are as close as possible to articulated aspiration vectors in terms of a given Lp-norm, may be derived using the Tchebychefftheory (of. [15]). (Correspondence with Professor A. Wierzbicki on this point is gratefully acknowledged.)

RALF OSTERMARKAND HANNU SALMELA

180 1.0 0.8 0.6 0.4 0.2 0.0

f

Dividends

Growth

\

Tax payments

/

Liquidity

ROI

Fig. 1. The graphicalconflictzone. The graphical conflict zone and an efficient solution derived by the system from an articulated aspiration profile is illustrated in Figure 1 above (cf. [17]). The underlying financial model was optimized separately with respect to each criterion over a planning period of five years, to produce the initial conflict zone. In the analysis the following criteria were used

(1) (2) (3) (4) (5)

Cumulative discounted dividend payments over the planning horizon. Growth as measured by the cumulative discounted turnover figure. Cumulative discounted tax payments. Liquidity as measured by the cash position over the planning horizon. Return on investment (ROI) over the planning horizon.

The content and definition of the criterion space is naturallyan intricate matter. From the viewpoint of the algorithm, however, these problems are secondary, as both independent and interdependent objectives can be specified. The graphical conflict zone shows at a glance the nature of conflicts between the criteria. Furthermore, by articulating aspiration profiles, the system may be used as a learning device for management in the process of determining suitable corporate strategies. At the same time, however, the issues of formulating the suitable criterion and, in particular, solution spaces as well as a detailed interpretation of the implications of alternative managerial actions are not captured by the system. A critical issue then is to design expert system support for these highly intellectual managerial tasks.

4. Examples of Rule Based Programming In this section, we cover some technical details of logic programming encountered in the design work of our pilot system. Specifically, the following topics will be discussed: (1) the decision-maker's perspective, (2) the database subsystem.

181

CONNECTING EXPERT SYSTEM FEATURES

We notice that the presentation is by no means representative of all problems of logic programming. However, we believe that the discussion serves to pinpoint some critical difficulties arising from trying to convert the logic of human thought into the logic of a knowledge system. As the pilot system was built with PROLOG, the discussion is based on the logic presupposed in this language. PROLOG is a rule based language, which also offers the advantages o f logic programming [11]. One of the distinctive features in using rule based programming is that the domain specific knowledge and the logic should have been built into the program [8].

4.1. PROGRAM LOGIC PERTAINING TO THE DECISION-MAKER'S PERSPECTIVE

At present, only three fundamental procedures are included: - keyboard entry of new problem description into new external case file, list existing problem description from case file onto screen, check existence of given activity in external data base. -

-

(i) Problem Description The approach is conventional, allowing the user to enter his problem description in an external file, as indicated in the flow chart given in Figure 2.

Problem - description ]

initial procedures: - set window frame - open datafile - set write device

I ' I enter description line I

I,,ne ="end" t

yes

no

l add new line I Fig. 2.

I set write device to screen mode

182

RALF OSTERMARK AND HANNU SALMELA

(ii) Problem Listing (Figure 3). Present - problem I I

initial procedures: set w i n d o w frame - open datafile set write device -

-

I output content of file on screen

I set device I to screen

I

end

I

Fig. 3. The mechanism for performing the file read procedure is shown below. Consider the following P R O L O G clauses: (1) fie_read(File in case)if readln(Line) and writeln(Line) and file_read(File in case). (2) file_read(_). In clause (1) a line in File in case is read and presented on the screen. As a second call to file_read( ) is encountered, the next line in the file is processed. Each time the file_read command is executed, two branches emerge: , file_read(File in case) false i , file_read(File_in_case) , true true file_read(casefile) , file_read(_) ~ exit true , file_read(_) The process continues as long as T R U E is returned for each tested branch. When the 'first FALSE value obtains (end of file), the program flow is directed from the previous (upper level) true-valued node to the other direction (exit).

183

CONNECTING EXPERT SYSTEM FEATURES

(iii) Check the status (allowed/forbidden) of a given activity (Figure 4) Activity Check ] Enter name of business Enter names of component business activities

I I Check status of entries against allowed set from external file

Fig. 4.

The status check consists of three logical procedures: (a) read list of activities at issue, (b) search the corresponding list of legally valid activities from an external file, (c) perform pairwise comparison between the sets in (a) and (b). The logical procedures are exemplified below. (a) read list of activities. Consider the following clauses: (1) read_list(HIT]) if readln(H) and not (H = "none") and ! and read_list(T). (2) read_list([ ]). Clauses (1) and (2) produce the following branch structure:

'

false ' read_list([HI T]) ~ read_list(T) true read_list(Case_activities)

read_list([H[ T])

, readlist([ ]) --, , exit

,

read_list([ ])

The process is terminated when the condition H = "none" is met. (b) Search activity file. The search activity continues until the item is found or end of file is reached, as shown in Figure 5.

184

R A L F O S T E R M A R K A N D H A N N U SALMELA Search activity file: - o p e n file - set read device

more_rules

+ ml read_term/allowed(Item,List))

+ m2 if n o t ( e o f ( F i l e )

-I

m21 read_term( )

a n d more_rules

I

I

m22 if not(eof(File)) and more rules

Fig. 5.

The final truth value of the search procedure is returned for subsequent processing. The search facility illustrates the typical program flow of the binary logic of PROLOG. (c) Pairwise comparison between the case list and the true (control) list. The testing is exemplified by the following P R O L O G subprograms: 1. 2. 3. 4. 5. 6.

member_list([ ],Truelist)!. member_list([ActivitypRest_list],True_list) if member(Activity,True_list) and member_list(Rest_list,Truelist). member(Activity,[Acitivityl_l)t. member(Activity,[_LRest_list]) if member(Activity,Rest_list).

Assume that the user gives an empty list to the comparison test subroutine. Then clause 1 will be true and no further testing is needed. Assume next that a nonempty list is entered. Then, clause 1 returns status FALSE, and program flow is directed to clause 2, where the nonempty list is split in two subsets, the first (Activity) containing the first element of the list, and the second (Rest_list) containing the rest. From clause 2 we proceed to clause 3, which directs the flow of control to clauses 5 and 6. If a match occurs at the first trial, clause 5 returns T R U E and control is returned to clause 4. Otherwise, clause 6 effects a looping through all elements in the control list. If no match occurs for the first element in the list, we are done and execution stops. Otherwise, the next element in the entered list is checked due to clause 4, which effects the above iteration through all elements as long as a match ultimately will occur.

CONNECTING EXPERT SYSTEM FEATURES

185

Model Builder/ 1 Analyst

I

Mathematical form search constraint search ]

database_check

existence I

no

I databaseadd I ~

Theoretical knowledge base format 1 (math format)

ask name I I

yes write item I

format 2 1 constraints Empirical knowledge base

l yes from I deletion _____~delete database no exit

I Fig. 6.

A natural extension of the DM's subsystem would be to add facilities for comparing and analyzing related problems and their solutions in the database. The presently known expert system shells do not readily provide tools for holistic interpretation of fuzzy problem situations nor their comparison to existing experience. By consequence, we have left these highly intellectual tasks outside the domain of an automized inference engine, to be tackled by human intellectual activity. 4.2. THE DATABASE SUBSYSTEM (Figure 6) The subsystem contains three connected functions: - existence checking of specified items with two alternative formats, - inclusion of new items in the database, - deletion of existing items from the database. The database is entered from either the model builder's or the analyst's subsystem. The database subsystem utilizes the _alternative domains facility of PROLOG. Consider the following domain and database declarations in P R O L O G format:

186

RALF OSTERMARK AND HANNU SALMELA

EXAMPLE DOMAINS T1,T2,T3 = symbol DATABASE_DOM = FI(T1,T2); F2(T3) DATABASE DATABASE_ITEM(DATABASE_DOM). In the example, the database contains two alternative formats for its records: F1 and F2. The former has two fields of type SYMBOL (T1, T2), whereas the latter has one SYMBOL typed field (T3). due to this construct, we may process completely different record structures without expanding the code unnecessarily for each new record format. 5. Results The present study was triggered by experience gained from the graphical decision support system designed and programmed by ()stermark [22] for commercial bank management. The conflict zone framework was accepted as a basis for the strategic planning system for the top management of Kansallis-Osake-Pankki, one of the largest commercial banks in Finland. The authors participated in the preliminary stages of system implementation, consisting of documentation and data collection from various parts of the organization and synthesization for managerial system use. Due to the complexity of the issues involved, for example, currency valuation and planning, forecasting key variables and consolidation of a multinational concern structure, a coherent formalization of the basic assumptions of strategic bank management proved to be difficult. The technical facilities of the algorithm provide a basis for evaluating consequences of alternative actions, rather than directly supporting an articulation of the fundamental assumptions, on which the analysis is carried out: A DSS- and MCDM-oriented planning device cannot provide intellectual guidelines for loading the system with relevant data and constraints. 6 Conclusion The general theme of the present study is thc usefulness of artificial intelligence in business economics. Our inquiry is structured on a case study involving a strategic planning systcm designed for bank environment. We feel, concordant with Holroyd et al. [4], that strategic policy analysis can be effectively supported with MCDM-techniques, perhaps even more effectively than with methods of rule-based inference. The use of optimization in strategic planning is not straightforward, howevcr. The unstructuredness of strategic planning requires that the system provide abilities to modify the problem definition used in the optimization process. In order to make

CONNECTING EXPERT SYSTEM FEATURES

187

the alterations, a decision-maker should have detailed knowledge on the actual model to be optimized and ability to comprehend the principles of the optimization algorithm.. This knowledge is normally expected from an analyst, not from a decisionmaker. As a possibility to aid the problem formulation process we propose the use of an expert system. We feel, however, that the expert system as such is not able to totally replace the human analyst. Thus, our proposal involves cooperation between the analyst and expert system in order to find a suitable structure for the problem to be solved. A clear-cut distinction between computerized versus human responsibility for the language system, the problem processing system and the knowledge system (see [5]) is not made in our system. According to Keim [6], the necessary knowledge components in AI are: factual, heuristic and meta knowledge, the last one denoting the concept of 'knowing what you know'. In our approach, the factual knowledge is stored in the actual model and in the problem definition to be used in optimization (decision variables, constraints, objectives etc.). The expert system is used to store heuristic knowledge and the human expert represents the meta knowledge. The central function of the expert system in problem structuring is the possibility to construct, store and retrieve conceptual frameworks. These frameworks facilitate communication between decision-maker, analyst and model builder. This function was also the main reason why rule based programming was adopted. Other functions could have equally well been constructed using data base facilities and a suitable query language. With logic programming and rule-based inference, it is relatively easy to present, on a conceptual level, managerial problems. Yet, difficulties in handling imprecision and uncertainty by means of logic programming have been emphasized in, for instance, [2, 6, 7, 10]. As these issues are central to business economics, an obvious direction for future efforts is to incorporate fuzziness and uncertainty within our AI framework. At the same time, however, the limitations of logic programming suggest a different approach, as proposed by, for example, [7] or [11]. References 1. Blanning, R. (1984) Knowledge acquisition and system validation ih expert systems for management, Human Systems Management 4, 280-285. 2. Carlsson, C. (1987) Fuzzy mathematical programming and approximate reasoning in integrated decision support and expert systems, working papers, Abo Academy 117/1987. 3. Hendry, L. C. (1987) The potential impact of artificial intelligence on the practice of OR, European J. Operational Res. 28, 218-225. 4. Holroyd, P., Mallory, G., Price, D. and Sharp, J. A. (1985) Developing expert systems for management applications, Omega 13, 1-11. 5. Holsapple, C. and Whinston, A. (1985) Management support artificial intelligence, Human Systems Management 5, 163-171. 6. Kasanen, E. and Ostermark, R. (1987) The managerial viewpoint in interactive programming with multiple objectives, Kybernetes 16, 235-240.

188

RALF tJSTERMARK AND HANNU SALMELA

7. Kasanen, E., Zeleny, M. and t3stermark, R. (1988) Gestalt system of holistic graphics: new management support view of MCDM. Discussion Paper to be presented at the Conference on MCDM in Manchester, August 1988. 8. Keim, R. T. and Jacobs, S. (1986) Expert systems: The DSS of the future? J. Systems Management (Dec.) 6-14. 9. Negoita, C. (1984) Management applications of expert systems, Human Systems Management 4, 275-279. 10. Uhr, L. (1985) Artificial intelligence: immediate promises and ultimate goals, Human Systems Management 5, 155-158. 11. Pollizer, E. and Jenkins, J. (1985) Expert knowledge, expert systems and commercial interests, Omega 13, 407-418. 12. Sheil, B. (1987) Thinking about artificial intelligence, Harvard Business Review (July-August), 91-96. 13. Steuer, R. (1986) Multiple Criteria Optimization: Theory, Computation and Application, John Wiley, New York. 14. Whalen, T. (1984) Fuzzy knowledge based systems in management, Human Systems Management 4, 262-274. 15. Wierzbicki, A. (1986) On the completenes and constructiveness of parametric characterizations to vector optimization problems, OR Spectrum 2, 73-88. 16. Ostermark, R. (1986) Designing financial model architectures. A study of managerial decision support. Publications of the Turku School of Economics, discussion papers 2: 1986. Presented at the DSS-87 Congress, San Francisco, 1987. 17. Ostermark, R. and Kasanen, E. (1985) A graphical decision support system for multi-objective financial modeling. Turku School of Economics, 1985. Presented at the International Conference on Operations Research, Lisbon 1986. 18. Ostermark, R. and Kasanen, E. (1988) Visualization of financial planning models: the case of an MCDM model in commercial banking. Discussion paper. 19. Ostermark, R. (1988a) Optimal compromising within a multicritical conflict zone, Operational Res. 35 255-262. 20. ~)stermark, R. (1988b) Aspiration profile preserving compromising within a multicriterial conflict zone, European J. Operational Res. 35, 263-270. 21. t3stermark, R. (1987) Star structure visualization of MCDM problems, unpublished working paper. 22. Ostermark, R. (1986) Linear goal and multiobjective programming in financial planning. Theoretical observations and a set of applications, Abo Academy, 1986.