Artificial intelligent systems architecture for strategic business decision making: a prototype neural expert system with what-if functions.

Lee, Kun Chang ; Han, Jae Ho ; Lee, C. Christopher 等

INTRODUCTION

Recently, a number of researchers in OR/MS (Operations Research/Management Science) have attempted to build intelligent expert systems for solving a wide variety of problems including production scheduling, finance, personnel, marketing, and accounting, etc (Waterman 1990). Common motivation underlying these researches is to intelligently assist decision-makers that have to solve poorly structured problems. The strategic planning problem is one of highly ill-structured OR/MS problems. Strategy, in effect, is the managerial action plan for achieving organizational objectives; it is mirrored in the pattern of moves and approaches devised by management to produce the desired performance. Strategy is therefore the how of pursuing the organization's mission and reaching target objectives (Thompson, Strickland III 1990). Today's managers have to think strategically about their company's position and about the impact of changing conditions. They have to monitor the external situation closely to know when the current strategy needs to be changed accordingly. The advantages of strategic thinking and conscious strategic planning activity include (1) providing better guidance to the entire organization on the crucial point of "what it is we are trying to achieve," (2) making management more alert to change, new opportunities, and threatening developments, (3) providing managers with a much-needed rationale that argues strongly for steering resources into strategy-supportive, results-producing areas, (4) helping unify the numerous strategy-related decisions by managers across the organization, and (5) creating a more proactive management posture and counteracting tendencies for decisions to be reactive and defensive. The advantage of being proactive versus reactive is that long-term performance is enhanced. Business history shows that high-performing enterprises often initiate and lead, not just react and defend. They see strategy as a tool for securing a sustainable competitive advantage and for pushing performance to superior levels.

Computer-based strategic planning systems are playing an increasingly relevant role in assisting both diagnosis of strategic problems likely to threaten the organization's performance and suggesting strategic alternatives to solve those problems. When designing such systems, certain objectives must be considered carefully. First of all, strategy analysts or managers in organizations should be able to use reliable, low-cost, user-friendly instruments--for example, programs running on personal computers. Nevertheless, to meet strategy analysts' requirements, processing time should be relatively short. Since any failure of such systems could prove seriously harmful to an organization's competitive position and performance, both fault tolerance and reliability are the crucial properties to be satisfied by such computer-based strategic planning systems. At the same time, the strategy analysts must be provided with as much information as possible about how the processing is carried out.

In an effort to accomplish these objectives, developers of computer aids for strategy analysts face a variety of problems deriving from the complex nature of strategic planning-related data. Such data is characterized by an intrinsic variability, which is the result of spontaneous internal mechanisms or as a reaction to occasional external stimuli. Furthermore, most events related to the strategic planning result from the interaction of many factors and sub-factors whose different effects are almost indistinguishable.

Strategy analysts are accustomed to such problems, but their skills cannot easily be incorporated into computer programs. Most strategic planning decisions are based on experience as well as on complex inferences and extensive strategic knowledge. Such experience and/or knowledge cannot be condensed into a small set of relations or rules, and this limits the performance of algorithmic approaches or conventional expert systems approaches to many strategic planning tasks. The breadth of strategic planning knowledge is therefore an obstacle to the creation of symbolic knowledge bases (for example, IF-THEN rules) comprehensive enough to cope with the diverse exceptions that occur frequently in practice. Experience-based learning, fault tolerance, graceful degradation, and signal enhancement are properties of neural networks that make the neural network-assisted expert systems effective in solving the strategic planning problems. This points to a way for implementing reliable computer-based strategic planning systems that can closely emulate a strategy analyst's expertise.

This paper presents the basic part of a prototype system named StratPlanner (Strategy Planner), which is a neural expert system for diagnosing strategic problems and suggesting strategic alternatives that seem appropriate for the current competitive situations. We will mainly focus on two issues: (1) the design of a neural expert system which is suitable for performing the "what-if" and/or "goal-seeking" analyses and (2) the competence of neural expert systems-driven strategic planning process in real strategic planning situations. Section 2 briefly discusses a basic theory of strategic planning and neural networks. Strategic planning techniques that are used in this paper are introduced in section 3. Inference mechanisms--forward inference and backward inference--are presented in section 4. In section 5, the performance of a prototype StratPlanner is illustrated with extensive experimental results. This paper is ended with concluding remarks in section 6.

STRATEGIC PLANNING AND NEURAL NETWORKS

A survey of the huge volume of contemporary practical and theoretical literature on neural network analysis yields the following three observations:

A There exists a great variety of viewpoints and approaches to neural network analysis.

B A general design principle that will help determine an appropriate architecture of neural networks for a particular application does not exist. It varies with the characteristics of applications.

C Major emphasis has been put upon experimental results obtained from extensive simulations, not upon rigorous theoretical derivations or proofs.

These general observations also prevail in neural network applications to OR/MS topics. The neural network analysis has embraced a very broad scope, from early success with neuron-like models called perceptrons (Rosenblatt 1961) and Adalines (Widrow, Hoff 1960) in the 1960's to the cooperative-competitive neural networks in the 1970's. Hopfield (1982) suggested an iterative computational neural network for associative retrieval and optimization, triggering a current explosion of interests in neural networks. Its theoretical basis was provided by many researchers. Some significant contributions have been made to the memory and learning models using competitive learning for autonomous feature extraction (Rumelhart, Zipser 1985) and delta learning for generalized information storage (McClelland, Rumelhart 1985). By extending these contributions, Rumelhart and his colleagues (Rumelhart, Hinto, Williams 1986) revived the backpropagation (sometimes called generalized delta) learning algorithm for the multilayer perceptrons, which has been successfully used in many experimental works (Lippmann 1987). Literature reporting the neural network applications to the OR/MS problems has recently begun to appear. White (White 1988) suggested a neural network analysis for economic prediction using the IBM daily stock returns data. Some neural network studies were performed to analyze a stock market prediction. Nonetheless, there exist a few studies that use neural networks for solving the strategic planning problems.

Neural networks have useful properties as follows (Gallant 1988, Zeidenberg 1990):

1 Generalization capability: When the training set contains noisy or inconsistent examples, during the learning phase the neural network can extract the hidden regularities residing in the set. After learning, the neural network can generalize, giving correct responses even in the presence of examples that are not included in the training set.

2 Graceful degradation: In addition, due to the neural network's noise rejection capability, performance is widely insensitive to noise corrupting the input patterns. In the presence of very noisy or contradictory inputs, neural network performance decays gradually.

3 Heuristic mapping: Furthermore, when there exists a kind of mapping function among the input-output pairs which is difficult to be represented by some statistical forms, the neural network tends to discover the mapping function in a very heuristic manner.

4 Fault tolerance: In addition, their parallel and distributed processing characteristics (information is spread throughout the neural network) make the neural networks widely insensitive to neurons (or processing units) and/or connection weights deficiencies or disconnections.

5 Multiple inputs: Finally, the neural networks can treat Boolean and continuous entities simultaneously. Therefore, despite the type (discrete or continuous) or source of input patterns, the neural networks can receive multiple kind of input patterns and deal with them effectively.

Because of all the properties mentioned above, the neural networks seem highly suitable for handling the strategic planning problems that are characterized by its unstructuredness and uncertainty.

STRATEGIC PLANNING TECHNIQUES

As is depicted in Figure 1, the process of strategic management consists of four basic elements: (1) environmental scanning, (2) strategy formulation, (3) strategy implementation, and (4) evaluation and control (Wheelen and Hunger, 1992). A number of strategic planning techniques have been proposed in previous researches (Abell, Hammond 1979, Glueck 1980, Larreche, Srinivasan 1982, Porter 1980, Rowe, Mason, Dickel 1982). Among them, the knowledge-based strategic planning approaches are well reviewed in Lee, Mockler and Dologite.

[FIGURE 1 OMITTED]

The available methods for strategic planning in the literature can be classified into three categories depending on their focuses: portfolio models, PIMS (Profit Impact of Market Strategy) analysis, and growth vector analysis. Refer to Lee (1992) for details about these three categories. Portfolio models assist managers in choosing the products that will comprise the portfolio and allocating limited resource to them in a rational way. The PIMS analysis is designed not only to detect strategic factors influencing profitability but also to predict the future trend of return on investment (ROI) in response to the changes in strategy and in market conditions. Growth vector analysis adopts the idea of product alternatives and market scope to support the product development strategy; this results in three strategies that are penetrating a market further with its present products, imitating competitors or introducing product variants, and innovating entirely new products.

We choose four strategic evaluation methods from portfolio models: BCG matrix, Growth/Gain matrix, GE matrix, and Product/Market Evolution Portfolio matrix. The reasons are that (1) portfolio models have been widely acknowledged among researchers and practitioners and (2) four strategic methods selected can provide most of the information that might have been expected from the PIMS analysis and growth vector analysis. The BCG matrix is the single most popular method. It emphasizes the importance of a firm's relative market share, industry's growth rate, and displays the position of each product in a two-dimensional matrix. The products are called "Stars", "Cash Cows", "Question Marks", or "Dogs" by the position in the BCG matrix as shown in Figure 2.

[FIGURE 2 OMITTED]

Usually the highest profit margins are expected from the "Stars", but they are also likely to require high net cash outflows in order to maintain their market shares. Eventually, the "Stars" will become "Cash Cows" as the growth slows down and the need for investment diminishes as it enters the maturity stage of the product life cycle. The "Question Marks" require large net cash outflows to increase the market share. If successful, these products will become new "Stars", which will in turn become the "Cash Cows" of the future. If unsuccessful, these products will become the "Dogs" to be excluded from the product portfolio. The BCG matrix alone is, however, not sufficient to make the investment decision because the model is too simple to cover the whole aspects of decision. In many circumstances, those factors other than relative market share and industry growth rate play a significant role in the production strategy formulation. To compensate the weakness of the BCG matrix, the Growth/Gain matrix, the GE matrix, and the Product/Market Evolution Portfolio matrix are used additionally.

The Growth/Gain matrix indicates the degree of growth of each product against the growth of market (See Figure 3). The product growth rate is plotted on the horizontal axis and the market growth rate on the vertical axis. Share gaining products appear below the diagonal line while share-losing products appear above it. The products on the diagonal line are interpreted as holding the current market share. Alternatively, the graph displaying the trends of the products sales compared with the market size may replace the role of Growth/Gain matrix in a simpler way (Lee 1985).

[FIGURE 3 OMITTED]

The composite measures of the market attractiveness and the business (product) strength are plotted in the GE matrix. In order to construct the GE matrix, managers have to select the relevant factors having significant relationship with the industry attractiveness and the business (product) strength of the firm. Next they assess the relative weights of those factors depending on manager's judgment, combining the weights to depict composite measures on the GE matrix. Figure 4 shows 3 x 3 GE matrix chart depicting relative investment opportunity.

[FIGURE 4 OMITTED]

Strategic managers can decide the overall direction of the firm through its corporate strategy by combining market attractiveness with the company's business strength/competitive position into a nine-cell matrix similar to the GE matrix. The resulting matrix, depicted in Figure 5, is used as a model to suggest some of the alternative corporate strategies that might fit the company's situation. Cell 1, 2, 5, 7, and 8 suggest growth strategies are either concentrated, which is expansion within the firm's current industry, or diversified, where growth is generated outside of the firm's current industry. Cells 4 and 5 represent stability strategies--a firm's choice to retain its current mission and objectives without any significant change in strategic direction. Cell 3, 6, and 9 display retrenchment strategies, which are the reduction in scope and magnitude of the firm's efforts.

[FIGURE 5 OMITTED]

GE matrix does not depict as effectively as it might the positions of new businesses that are just starting to grow in new industries. So, in that cases, Hofer and Schendel [9] proposed to use a Product/Market Evolution matrix in which businesses are plotted in terms of their relative competitive position and their stage of product/market evolution. They also recommended investment strategies at the business level (See Figure 6).

The combined use of these four strategic models can provide most of the functions necessary to effectively evaluate the corporate and/or business strategies.

[FIGURE 6 OMITTED]

INFERENCE MECHANISM

The multi-phased aspects of strategic planning activities described above indicate that one-shot or wholesome approach is not appropriate for an effective strategic planning. Rather, to simulate a strategy analyst's reasoning as closely as possible, it would be better to divide the strategic planning-related decision-making processes into a relevant small number of subprocesses. In this respect, a forward inference mechanism suggests more robust strategies. Forward inference process helps decision-makers perform a "what-if" analysis that is essential for diagnosing the strategic problems and preparing strategic policies against the uncertain future.

[FIGURE 7 OMITTED]

The forward inference process is composed of two stages. The first stage uses a relative competitive position (RCP) neural network module, which suggests the competitive positions in a target market. The second stage uses both a generic business strategy (GBS) neural network module and a contingency corporate strategy (CCS) neural network module. Each module consists of one feed-forward neural network trained by the backpropagation algorithm. Also, the stage of market evolution (SME) and industry attractiveness (IA) are also used as additional information to the CCS and GBS neural network module, as shown in Figure 8.

[FIGURE 8 OMITTED]

In the first stage, the RCP neural network module provides information about the competitive position in the market relative to that of a target competitor. We considered two kinds of strategic planning models: BCG and Growth/Gain matrix. The architecture of the RCP neural network module has 22 neurons in the input layer and 4 neurons in the output layer (See List 1). For comparing relative competitive position between non-leading firms at the specific market, we modified the number of BCG matrix's cells from 4 to 8. The output value derived from this neural network module is used as the input value of RCP part of the CCS and GBS neural network modules. Following is a list of RCP neural network module architecture.

List 1

RCP Neural Network Module

Input Neurons:

 < Decision Making Company >
 BCG : Stars/Cash Cows/ H-Question Marks/ M-Question Marks/
 L-Question Marks/ H-Dogs/ M-Dogs/ L-Dogs
 Growth/Gain: Share Gainer/ Share Holder/ Share Loser
 < Competitor >
 BCG : Stars/ Cash Cows/ H-Question Marks/ M-Question Marks/
 L-Question Marks/ H-Dogs/ M-Dogs/ L-Dogs
 Growth/Gain: Share Gainer/ Share Holder/ Share Loser
Output Neurons:
Strong/ Average/ Weak/ Drop-Out

In the second stage, a choice is made between the GE and the Product/Market Evolution matrices according to the nature of the company's business. The criterion recommended by Hofer and Schendel (Hofer, Schendel 1978) is that if most of the businesses represent aggregations of several product/market segments, the GE matrix is more suitable; and if most businesses consist of individual or small groups of related product/market segments, a Product/Market Evolution matrix should be used. If the decision maker has difficulty making a decision based on these considerations, he should use both types of matrices to see which fits more appropriately to his own situation.

IA presents information about industry attractiveness. In this paper, to determine the degree of industry attractiveness being considered, decision-maker is prompted to select appropriate criteria, and determine their weights and ratings in five scales. According to the sum of weighted scores, one of 4 areas (High, Medium-High, Medium-Low, Low) is presented. Combining the results from RCP module and IA module, the CCS neural network module provides one of ten cells of GE matrix. The architecture of CCS neural network module is summarized in List 2.

SME presents information about the stage of the market for a product development stage, growth stage, shakeout stage, maturity stage, and decline stage. To determine an appropriate market stage of a product being considered in this paper, decision-maker is prompted to select one of the five stages.

List 2

CCS (Contingency Corporate Strategy) Neural Network Module

Input Neurons:

 RCP Part : Strong/ Average/ Weak/ Drop-Out
 IA Part : High/ Medium-High/ Medium-Low/ Low

Output Neurons :

H-S Winners/ H-A Winners/ M-S Winners/ H-Average
Businesses/ L-Average Businesses/ Profit Producers/ Question
Marks/ M-W Losers/ L-A Losers/ L-W Losers

Combining the RCP neural network module with SME information, GBS neural network module provides one of six types of generic business strategies that follow: share increasing strategy, growth strategy, profit strategy, market concentration/ asset reduction strategy, liquidation strategy, and turnaround strategy. The architecture of GBS neural network module is shown in List 3.

List 3

GBS Neural Network Module

Input Neurons :

 RCP Part
 Strong/ Average/ Weak/ Drop-Out
 SME Part
 Development/ Growth/ Shake-Out/ Maturity/ Decline

Output Neurons :

 Share Increasing/ Growth/ Profit/ Market Concentration and
 Asset Reduction/Turnaround/ Liquidation or Divestiture

After training the RCP, CCS and GBS neural network modules with appropriate training data, three sets of neural network knowledge bases are generated. They are RCP knowledge base, CCS knowledge base, and GBS knowledge base.

Expert's knowledge is stored in a conventional knowledge base that may include information about various topics, for example, industry environments, socio-economic situations, contingency corporate strategies, competitive position objective, and investment strategy with respect to various strategic situations, etc. Especially, expert knowledge related to three kinds of areas--contingency corporate strategies, competitive position objective, and investment strategy--are considered. Contingency corporate strategies include nine types of strategies: "concentration via vertical integration", "concentration via horizontal integration", "concentric diversification", "conglomerate diversification", "pause or proceed with caution", "no change in profit strategy", "turnaround", "captive company or divestment," and "bankruptcy or liquidation". Each of the six generic types of business strategies involves a different pattern of competitive position objectives, investment strategies, and competitive advantages, which are summarized in Table 1.

ILLUSTRATION

Architecture of StratPlanner

We developed a prototype StratPlanner running on Windows 2000. It is coded in Microsoft Visual C++ language. Its main menu is composed of five sub-menus as shown in Figures 9 and 10.

[FIGURE 9 OMITTED]

[FIGURE 10 OMITTED]

As mentioned in introduction, we will illustrate the performance of forward inference mechanisms: what-if analysis. For example, in StratPlanner, what-if analysis is performed in accordance with the steps shown in List 4.

Figure 11 illustrates showing the result from CCS neural network knowledge base.

List 4

Steps of StratPlanner associated with What-If analysis.

Stage 1: RCP Stage

 Step 1. Open the weight file of RCP neural network module.
 Step 2. Select a target product.
 Step 3. Input data about BCG and Growth/Gain matrix.
 Step 4. Get the result from RCP neural network knowledge
 base.

List 4

Steps of StratPlanner associated with What-If analysis.

Stage 2: GBS and/or CCS Stage
 If GBS analysis is selected, then perform the following
 steps.
 Step 1. Open the weight file of GBS neural network
 module.
 Step 2. Input data about stage of market evolution.
 Step 3. Get the result from GBS neural network
 knowledge base.
 If CCS analysis is selected, then perform the following
 steps.
 Step 1. Open the weight file of CCS neural network
 module.
 Step 2. Input data about industry attractiveness.
 Step 3. Get the result from CCS neural network
 knowledge base.

[FIGURE 11 OMITTED]

Data

Experiments were performed with Korean automobile data, which is fabricated as a strategically turbulent market designed to show the performance of StratPlanner in a turbulent strategic planning environment. Table 2 shows the categories of automobile data used in our experiments.

Monthly domestic sales data of three companies' passenger cars from May 1990 to August 1994 as well as miscellaneous strategic planning data from May 1990 to August 1994 was collected. The domain knowledge from two experts, a strategy analyst in 'K' automobile company and a strategy expert in university was also used in this experiment. Table 3 shows the type and description of data used in our experiments.

The data set consisted of 52 cases divided into 32 cases from May 1990 to December 1992 for the training set and 20 cases from January 1993 to August 1994 for the test set. Another data set is arranged for the differences in production periods. Based on this data, we trained and tested RCP, CCS, GBS, CCS_RCP, GBS_RCP, RMS_GG neural network modules. By using monthly data, this experiment is assumed to be a monthly one-shot.

Experiment for Forward Inference

For illustration of forward inference, consider KIA as a decision-making company. Suppose that KIA wants to build two kinds of strategies for its small type car "PRIDE" using data of Jan. of 1993: (1) competitive position strategy and (2) investment strategy. Analysis of the current period's data represents that current competitive positions of "PRIDE" compared to its major competitor, HYUNDAI's "EXCEL", are "High-Dogs" and "Share Loser", respectively. Similarly, the competitive position of HYUNDAI's "EXCEL" is analyzed to belong to "Cash Cows" in the BCG matrix and "Share Gainer" in the Growth/Gain matrix, respectively. Using this information, the RCP neural network knowledge base presents a "Weak" position. The stage of small car market evolution is analyzed as "Maturity". Based on the results from RCP neural network knowledge base and the stage of market evolution, the GBS neural network knowledge base provides "Market Concentration/Asset Reduction" strategy. The sample screen of this result is shown in Figure 12. This process by RCP and GBS neural network modules and other test cases are summarized in Table 4.

This generic business strategy is inputted to the conventional knowledge base, firing the following two rules.

IF Generic_Business_Strategy = Market_Concentration/Asset_Reduction

THEN Competitive_Position_Objective = "Reduce position to smaller defensible position"

IF Competitive_Position_Objective = "Reduce position to smaller defensible position"

THEN Investment_Strategy = "Moderate to negative investment"

DISPLAY "Usually some new assets are required, while others are sold off. The net level of investment depends upon the relative proportion of these two activities in each specific case"

Figure 13 depicts the result of forward inference. In response to the current market situations of KIA's PRIDE, StratPlanner provides "Reduce position to smaller defensible position" strategy as a competitive position objective and "Moderate to negative investment" strategy as an investment strategy.

[FIGURE 12 OMITTED]

[FIGURE 13 OMITTED]

CONCLUDING REMARKS

In this paper, we proposed a neural expert system capable of performing a forward inference so that strategic planning problems may be solved more effectively. The proposed neural expert system is designed to provide "what-if" inference function, based on combining the generalization capability of neural networks with expert system. A prototype system StratPlanner was proposed to prove our approach. Its performance was illustrated with real competitive data of Korea Automobile Industry. However, there exist much room for further research. First, "goal-seeking" function can be added to a future system development to make the system capable of performing a bi-directional inference. Goal seeking functions are realized through the backward inference mechanism, enabling the neural expert system to show the appropriate inputs (or conditions) to guarantee the desired level of outputs. Second, an improved version of StratPlanner can incorporate refined mechanisms of environmental analysis, competitor analysis, and advanced strategic planning models.

REFERENCES

Abell, D.F. & J.S. Hammond (1979). Strategic market planning: problems and analytical approach. Prentice-Hall, Englewood Cliffs, NJ.

Ackley, D.H., Hinton, G.E. & T.J. Sejnowski (1985). A learning algorithm for Boltzmann machines. Cognitive Science 9, 147-169.

Clark, Delwys N & Scott, John L (1999). Strategic level MS/OR tool usage in the United Kingdom and New Zealand: A comparative survey. Asia--Pacific Journal of Operational Research; Singapore, 16(1), 35-51.

Cohen, M.A. & S. Grossberg (Sept., 1983). Absolute stability of global pattern formation and parallel memory storage by competitive neural networks. IEEE Transactions on Systems, Man, and Cybernetics, SMC 13, 815-825.

Fukushima, K., Neocognitron (Sept., 1983). A self-organizing neural network model for a mechanism of pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI 8, 398-404.

Gallant, S.I., (Feb. 1988). Connectionist expert system. Communications of the Acm, 31(2), 152-169.

Glueck, W.F., (1980). Business policy and strategic management, McGraw-Hill, Grossberg, S., (1976). Adaptive pattern classification and universal recording: Part 1. Parallel development and coding of neural feature detectors. Biological Cybernetics, 23, 121-134.

Grossberg, S., (1982). Studies of mind and brain, Reidel, Hingham, Mass Hatchuel, Armand; Liberatore, Matthew J; Weil, Benoit; & Stylianou, Antonis C (2000). An organizational change perspective on the value of modeling. European Journal of Operational Research; Amsterdam, 125(1), 184-189.

Hofer, C.W. & D. Schendel, (1978). Strategy formulation: analytical concepts, West Publishing Co.,

Hopfield, J.J., 1982. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences USA, National Academy of Sciences, Washington, D.C., 79, 2554-2558.

Honavar, Vasant & Uhr, Leonard (1994). Artificial Intelligence and Neural Networks: Steps Toward Principled Integration. Boston: Academic Press.

Honavar, Vasant & Uhr, Leonard (1984). Neurons with graded response have collective computational properties like those of two-state neurons. Proceedings of the National Academy of Sciences USA, 81, 3088-3092.

Honavar, Vasant & Uhr, Leonard (1984). & D.W. Tank, Neural computation of decision in optimization problems. Biological Cybernetics, 52, 1985, 141-152.

Jacobs, R.A., (1988). Increased rates of convergence through learning rate adaptation. Neural Networks, 1, 285-307.

Kamijo, K. & T. Tanigawa, (1990). Stock price pattern recognition: A recurrent neural network approach, Int'l Joint Conference on Neural Networks, San Diego, Calif., 1, 215-221.

Kang, Boo Sik; Choe, Deok Hyoen & Park, Sang Chan (1999). Intelligent process control in manufacturing industry with sequential processes. International Journal of Production Economics; Amsterdam, 60-61, 583-590.

Kimoto, T. & K. Asakawa, (1990). Stock market prediction system with modular neural networks. Int'l Joint Conference on Neural Networks, San Diego, Calif., 1, 1-6.

Kohonen, T., (1987). State of the art in neural computing. Int'l Joint Conference on Neural Networks, San Diego, Calif., 1, 79-90.

Larreche, J.C. & V. Srinivasan, (1982). STRATPORT: A Model for the Evolution and Formulation of Business Portfolio Strategies. Management Science 28(9), 979-1001.

Lederer, Albert L & Sethi, Vijay (1998). Seven guidelines for strategic information systems planning. Information Strategy; Pennsauken, 15(1), 23-28.

Lee, H.K., (1985). Interactions of Long-term Planning and Short-term Planning: An Intelligent DSS by Post Model Analysis Approach, Unpublished Master Thesis, Dept. of Management Science, Korea Advanced Institute of Science and Technology.

Lee, K.C., (1992). Synergism of Knowledge-based Decision Support Systems and Neural Networks to Design an Intelligent Strategic Planning System. Journal of the MIS Research (in Korea), 35-56.

Li, Heng & Love, Peter E D (1999). Combining rule-based expert systems and artificial neural networks for mark-up estimation. Construction Management and Economics; London, 17(2), 169-176.

Li, Shuliang (2000). The development of a hybrid intelligent system for developing marketing strategy. Decision Support Systems; Amsterdam, 27(4).

Lippmann, R.P., (1992, April, 1987). An introduction to computing with neural nets. IEEE ASSP Magazine, 3, 4-22.

McClelland, J.L. & D.E. Rumelhart, (1985). Distributed memory and the representation of general and specific information. Journal of Expert Psychology: General, 114, 158-188.

McClelland, J.L. & D.E. Rumelhart, (1986), Parallel distributed processing: explorations in the microstructures of cognition, Vol. 2, MIT Press, Cambridge, Mass.

Mingers, J (2000). The contribution of critical realism as an underpinning philosophy for OR/MS and systems. The Journal of the Operational Research Society; Oxford, 51(11), 1256-1270.

Mockler, R.J., & D.G. Dologite, (1991). Knowledge-based systems to support strategic planning decision making. Proceedings of the Twenty-Fourth Annual Hawaii International Conference on System Sciences, 3, 173-180.

Moriarty, Stephen (1998). Fixed assets management: Beyond the spreadsheet. Management Accounting; London, 76(8), 42-44.

Porter, M.E., (1980). Competitive strategy: techniques for analyzing industries and competition, MacMillan, NY.

Rabiner, L.R & B.H. Juang, (Jan., 1986). An introduction to hidden Markov models. IEEE ASSP Magazine, 3, 4-16.

Ray, Tsaih; Hsu, Yenshan & Lai, Charles C (1998). Forecasting S&P 500 stock index futures with a hybrid AI system. Decision Support Systems; Amsterdam, 23(2), 161-174.

Rosenblatt, F., (1982). Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Spartan, Washington, D.C [27] Rowe, A.J., R.O. Mason & K. Dickel, Strategic management & business policy: a methodological approach, Addison-Wesley, Reading, Massachusetts.

Rumelhart, D.E., G.E. Hinton, & R.J. Williams, (1986). Learning internal representations by error propagation. Parallel Distributed Processing (PDP): Exploration in the Microstructure of Cognition Vol. I, Cambridge, MA:MIT Press, ch.8, 318-362.

Rumelhart, D.E. & D. Zipser, (1985). Feature discovery by competitive learning. Cognitive Science, 9, 75-112.

Steinbush, K., (1961). The learning matrix, Cybernetics, 36-45.

Thompson, A.Jr. & A.J. Strickland III, (1990). Strategic management: concepts and cases, Irwin, 5th Edition.

Vandaele, Nico J; Lambrecht, Marc R; DeSchuyter, Nicolas; & Cremmery, Rony (2000). Spicer Off-Highway Products Division--Brugge improves its lead-time and scheduling performance. Interfaces: Providence, 30(1), 83-95.

Waalewijn, P., & R.H. Boulan, (1990). Strategic planning on a personal computer. Long Range Planning, 23(4), 97-103.

Walczak, Steven & Cerpa, Narciso (1999). Heuristic principles for the design of artificial neural networks. Information and Software Technology; Amsterdam, 41(2),107-117.

Waterman, D.A., (1986). A guide to expert systems, Addison-Wesley Publishing Co.

Wheelen, T.L. & J.D., (1992). Hunger, Strategic management and business policy, Addison-Wesley Publishing Co.

White, H., (1988). Economic prediction using neural networks: The case of IBM daily stock returns. Int'l Joint Conference on Neural Networks, San Diego, Calif., 2, 451-458.

Widrow, G. & M.E. Hoff, (1960). Adaptive switching circuits. Institute of Radio Engineers, Western Electronic Show and Convention, Convention Record Part 4, 96-104.

Willshaw, D.J., (1971). Models of distributed associative memory models, Ph.D. Dissertation, University of Edinburgh, Scotland.

Zeidenberg, M., (1990). Neural network models in artificial intelligence, Ellis Homewood, England.

Kun Chang Lee, Sung Kyun Kwan University

Jae Ho Han, Pukyong National University

C. Christopher Lee, Central Washington University

Table 1
Characteristics of the Six Generic Business Strategies
(Hofer & Schendel, 1978)

 Type of Generic Competitive Position Investment Strategy
 Strategy Objective

 Share-increasing
 strategies

 Development stage Increase position Moderate investment

 Shake-out stage Increase position High investment

 Other stages Increase position Very high investment

 Growth strategies Maintain position High investment

 Profit strategies Maintain position Moderate investment

 Market concentration Reduce (shift) Moderate to negative
 and asset reduction position to smaller investment
 strategies defendable level
 (niche)

 Liquidation or Decrease position to Negative investment
divestiture strategies zero

Turnaround strategies Improve positions Little to moderate
 investment

Table 2
Categories of Korean Automobile Data

 Company

Type Car KIA HYUNDAI DAEWOO

Small Pride Excel Lemans
Compact Capital Sephia Elantra Espero
Medium Concord Sonata Prince
Large Potentia Grandeur Super Salon

Table 3
Type and Description of Data Used in Experiments

 Type of Data Description of Data

Quantitative Monthly Sale Data Market Growth Rate
 Relative Market Share
 Product Growth Rate

Qualitative Expert Knowledge Preparation of Input/Output Pairs
 Data Supervised Learning
 Preparation of Desired Output used
 in Test
 Knowledge related to three kinds
 of areas:
 Contingency corporate strategies
 Competitive position objective
 by type of generic strategy
 Investment strategies by type of
 competitive position objective

 Data Produced by Relative Competitive Position
 Neural Network Position in GE Matrix
 Modules Position in Product/Market Portfolio
 Matrix
 Position in BCG Matrix
 Position in G/G Matrix

Table 3
Type and Description of Data Used in Experiments

Type of Data Description of Data

User's Determination of stage of market by car type
Judgement Determination of industry attractiveness by car type
 Variable Selection
 Weight Determination

Table 4: Illustration of forward inferencing by RCP and GBS
neural network modules

Test Set Decision Making Company Competitor
 (KIA's PRIDE) (HYUNDAI's EXCEL)

 BCG G/G BCG G/G

 93.01 High-Dogs Share Loser Cash Cows Share Gainer

 02 High-Dogs Share Gainer Cash Cows Share Holder

 03 High-Dogs Share Loser Cash Cows Share Gainer

 04 High-Dogs Share Gainer Cash Cows Share Gainer

 05 Cash Cows Share Gainer High-Dogs Share Loser

 06 Cash Cows Share Loser High-Dogs Share Holder

 07 Cash Cows Share Gainer High-Dogs Share Loser

 08 High-Dogs Share Loser Cash Cows Share Gainer

 09 High-Dogs Share Gainer Cash Cows Share Loser

 10 High-Dogs Share Loser Cash Cows Share Holder

 11 Middle-Dogs Share Loser Cash Cows Share Loser

 12 Middle-Dogs Share Loser Cash Cows Share Loser

 94.01 Middle-Dogs Share Loser Cash Cows Share Loser

 02 High-Dogs Share Gainer High-Dogs Share Loser

 03 Middle-QM Share Loser Stars Share Gainer

 04 Stars Share Holder Low-Question Share Loser
 Marks

 05 Cash Cows Share Loser Low-Dogs Share Loser

 06 Low-QM Share Loser Low-Question Share Loser
 Marks

 07 Cash Cows Share Gainer Low-Dogs Share Loser

 08 Low-Dogs Share Loser Low-Dogs Share Loser

Test Set RCP SME GBS

 Actual Desired

 93.01 Weak Maturity Market Market
 Concentration Concentration

 02 Weak Maturity Market Market
 Concentration Concentration

 03 Weak Maturity Market Market
 Concentration Concentration

 04 Weak Maturity Market Market
 Concentration Concentration

 05 Strong Maturity Profit Profit
 Strategies Strategies

 06 Strong Maturity Profit Profit
 Strategies Strategies

 07 Strong Maturity Profit Profit
 Strategies Strategies

 08 Weak Maturity Market Market
 Concentration Concentration

 09 Weak Maturity Market Market
 Concentration Concentration

 10 Weak Maturity Market Market
 Concentration Concentration

 11 Weak Maturity Market Market
 Concentration Concentration

 12 Weak Maturity Market Market
 Concentration Concentration

 94.01 Weak Maturity Market Market
 Concentration Concentration

 02 Strong Maturity Profit Profit
 Strategies Strategies

 03 Weak Maturity Market Market
 Concentration Concentration

 04 Strong Maturity Profit Profit
 Strategies Strategies

 05 Strong Maturity Profit Profit
 Strategies Strategies

 06 rop-out Maturity Liquidation Profit
 or Divestiture Strategies
 Strategies

 07 Strong Maturity Profit Profit
 Strategies Strategies

 08 Strong Maturity Profit Profit
 Strategies Strategies