Key Words: genetic algorithms, Baldwin Effect, decision support systems, human computer interface (HCI) design, graphical user interface (GUI) design

Abstract

Introduction

In the past fifteen years, various computer-based tools and models have been developed to support the decision-making process in business and industry. Expert systems, neural networks, data repositories, and data management systems have emerged as likely candidates for computer-based decision support products (Caster, 1994). Though these tools have proven effective in some limited environments, none has been flexible and reliable enough to support decision making in highly complex or uncertain conditions (Wyatt, 1995). All of these decision support tools are based on conceptual designs that assume answers to two fundamental questions: What is the decision-making process? and How can a computer system support that process most effectively? This paper examines both of these questions, with the goal of defining a conceptual model for a learning genetic algorithm to be used as an effective decision support system.

Two models of decision making are described: Limited rationality and rule following. Recent research findings related to each model are discussed. Next, these decision-making models are related to the fundamental assumptions for the conceptual design of computer-based decision support systems. Major types of existing decision support systems are described in terms of the decision-making models on which they depend. A genetic algorithm (GA) model is investigated as a model of organizational change (Bruderer & Singh, 1996). Finally, the same GA is evaluated as a possible engine for a decision support system, and requirements and a conceptual design of such a decision support system are proposed.

Decision Making

Limited Rationality. Until the early 1970s, only one model--the model of rational choice--was widely accepted (Cohen, March, & Olsen, 1972). This model is procedural in nature. It depends on a logical understanding and analysis of consequences that will follow from the individual's decision. The decision maker is expected to be aware of all reasonable alternatives, the expected consequences of each alternative, his or her own preferences with regard to the consequences, and a rule that allows the decision maker to select among alternatives, based on available information. Each individual is expected to make a choice that maximizes his or her own utility function. This model was widely held by investigators because it provided extensive reliability and predictive power, given initial conditions and stated preferences (Anderson, 1995; Morecroft, 1983; Simon, 1979).

Slovic, 1995), and relative recency or salience of an event (Akerlof, 1991) are among the factors that have been proven to distort the expected rationality of decisions. All of these studies raise questions about the validity of the theory of limited rationality in decision making.

Theories of rational choice, though reliable and internally consistent, did not always describe decision behavior sufficiently. The validity of the theory was challenged in many contexts. Over time, adaptations were made in the strict rational choice model to incorporate calculations of risk and ambiguity (Tversky & Fox, 1995). Risk is characterized as the probability that something will or will not happen. This introduces a probabilistic component to the rationality of choice. The decision maker might not know exactly what will happen, but have an idea of the probability of consequences from each rational alternative. With this information, a rational and predictable decision could follow. Ambiguity, on the other hand, introduces the concept of uncertainty in decision making. The decision maker might not know all of the alternatives at his or her disposal. The consequences of one or more of the alternatives might be unknown. Other persons' decisions, unknown to the decision maker, might influence the outcome of the rational decision. Various models have incorporated methods of dealing with risk and ambiguity within the constraints of rationality (Evans & Over, 1996; Fox, 1995; Frisch & Clemen, 1994; Hollenbeck et al., 1994; Morecroft, 1983; Simon, 1982; Weber, 1994). These adjusted versions of rational choice theory have come to be known as theories of bounded or limited rationality.

The central concept of preference provides another challenge to rational choice. Within models of limited rationality, a subject's preferences are considered to account for variation across a population. Preference is the internal judgement process that ultimately determines the decision maker's choice. The decision maker may or may not be aware of an immediate preference. Preferences change over time, though they are assumed to be consistent across decisions made at the same time. In short, preference becomes the factor that adjusts all other factors to generate the decision maker's final, rational choice. However irrational a decision may appear, the observer can assume that it was made rational through the preference of the decision maker (Anderson & Block, 1995).

Individuals' preferences vary over time providing another indication that rational choice models are insufficient to account for behavior of decision makers.

" . . . preference reversals violate the principle of procedure invariance that is fundamental to theories of rational choice and raise difficult questions about the nature of human values. If different elicitation procedures produce different orderings of options, how can preferences be defined and in what sense do they exist? Describing and explaining such failures of invariance will require choice models of far greater complexity than the traditional models." (Slovic, 1995, 1)

My personal experience confirms the findings of the formal research cited above. As an executive officer of a corporation, I am intimately involved in the decision-making process every day. I expect my staff and myself to make rational decisions based on perfect information and careful analysis, but my observations prove otherwise. Seldom is there sufficient information, time, or insight to make a reliable prediction about the consequences of our choices. Our explanations and justifications of our decision-making process sound perfectly rational, but the reality is quite different.

Both research and experience indicate that the rational choice approach to decision making is insufficient to explain either the experience or the outcomes of the decision process. A more effective model is required to capture the complexity of the process.

Rule Following. The second category of decision-making models involves rule following (Burns & Flam, 1987). In this model, each decision maker has a variety of rules that have evolved for him or her over time. The accumulation of those rules establishes the decision maker's identity. In this model, when the decision maker is asked to choose, he or she asks three questions:
What kind of situation is this? (Recognition) What kind of person am I? (Identity) What does a person such as I do in a situation like this? (Rules) (March, 1994)

Rather than being procedural, these models are evolutionary in nature. The three stages that generate and test emergent rules and identities in the problem solving process can be compared with the stages of evolutionary change: variation, selection, and retention. (March, 1994; Weick, 1976). The decision maker investigates the requirements of the environment, determines which applications of which rules will be most adaptive in response, then tests the adaptation against reality. The fittest rules survive to construct the individual identity, and the fittest identities survive to determine the most effective decision makers.

Persons who investigate the behavior of complex adaptive systems (CAS) have applied their theoretical structures to explain decision making and change processes in organizations (Bruderer & Singh, 1996; Forrester, 1994; Gell-Mann, 1994; Holland, 1994; Kauffman, 1995; ). They use the metaphor of emergence in CAS to represent the evolving nature of decision making in a rule-following system. These perspectives of adaptation and fit find parallels in economic and psychological studies of decision-making processes.

In 1972, Cohen, March, and Olsen introduced a complex and dynamical image of the decision- making process--the garbage can. This model represents decisions that emerge from complex interactions among actors, solutions, problems, and choice opportunities. The decision-making process in such a complex environment involves sorting in temporal order. When events, actors, solutions and so on are remote in time, they are seen as remote in connection and mutual influence. On the other hand, when they are temporally local, they are considered to be causally linked (March, 1994). In its conceptual content, this model appears to be a special case of the fitness or rule-following models. In the garbage can, the temporal variable is used as one-dimensional fitness measure.

The exploration and exploitation model (March, 1991) is another variation of the evolutionary theme. It describes an organization shifting between two basic activities. Exploration involves random variation and selection to determine strategies that fit the environment. Exploitation involves focused repetition of the fittest behavior in the stage of retention. These phases of behavior alternate, giving the same on-going patterns as the evolutionary cycle of decision making.

increased.

Other researchers also use the model of rule and fitness to investigate decision-making behavior. Like other evolutionary changes, outcomes are not predictable. Random variation shapes the options from which the fittest strategy is selected. In decisions, equivalent processes of elicitation (or emergence) give rise to systematically different responses (Slovic, 1995). Rational choices are frequently seen to adapt to environmental forces. Comparative ignorance (Fox & Tversky, 1995) and relative similarity among options (Mellers & Biagini, 1994) both influence choices among givens. Nonlinearity has been recognized in both risky and uncertain decisions. Decision makers tend to overweigh low probabilities and underweigh mid-level and high probabilities. In other words, they are more likely to be swayed by shifts from the impossible to the possible or from the possible to the probable than they will be by changes from less to more possible or less to more probable (Tversky & Fox, 1995). Though Anderson and Block (1995) challenged his findings, Akerlof (1991) found that salience or recency gives greater weight to information or experience that is more recent or in the near future, rather than temporally remote events. All of these behaviors provide for patterns and discontinuities that one might expect from a complex adaptive or evolutionary system (Bak, 1996; Dooley, 1997; Guastello, 1995).

The decision-making conversations in which I am involved in my profession reflect these complex and nonlinear patterns. Many dimensions of expectation and demand are discussed that affect the decision (customer expectations, cost, employee satisfaction, personal relationship, reputation in the community, anticipated marketing opportunities, etc.). Historical patterns of decisions made and their consequences are assessed to discover meaningful precedents. Possible solutions are brainstormed and each is compared and contrasted to historical, existing, and anticipated patterns. The solution that "fits" these complex and emerging criteria is the one that is implemented.

All of these formal studies and my personal experience indicate that a rule-following model of decision making includes all three steps of evolutionary process: variation in rules and identities based on context, adaptation of rules and identities to move closer to a goal, and selection of decision strategies and outcomes that prove the best fit for the problem space. A computer-based tool that will support a rule-based decision-making process must model evolving fitness toward a goal.

Decision Support Tools

Data repositories and data management systems seek to provide information about the decision space to reduce the limitation or boundedness of the rational decision maker. The assumption is that the decision maker will incorporate all new data to make his or her image of the present and the future more accurate. As a result, the decision will be improved. Expert systems capture the imminently rational decisions made by the most competent experts. An expert system allows the user to reproduce the reliable decision by "borrowing" the rational insight of others. Procedural analysis tools guide the decision maker through analytical steps that are intended to reduce uncertainty, increase information, and encourage competent and rational decisions. A neural net must be trained in reasonable alternatives, the expected consequences of each alternative, and preferences with regard to the consequences. The process of training a neural net may be evolutionary and adaptive, but in use it serves as a surrogate rational decision maker.

System dynamics models have been used successfully in some environments to support the decision-making process (Morecroft & Sterman, 1994; Roberts et al., 1983). A simulation model is built for the environment, and decision makers can use the model to anticipate outcomes of specific decisions. As a support tool, these systems seek to make the connections between decision and outcome predictable, as they should be in a rational choice model. In so far as the model represents reality, these tools are excellent support for real managers. The excessive resources and expertise required to construct and maintain such models, however, make them impractical in most decision-making situations.

If it is true that decisions are made through a process of random variation, adaptation, and selection, then an effective decision support tool will represent this process and provide assistance to the decision maker as he or she moves through the process. Both the garbage can and the exploration/exploitation models have been used to design computer-based simulation models of decision making, but they have not been used to create tools to support the decision-making process as such. A genetic algorithm will be an appropriate tool to support business decision making because it will:
Simulate the process of variation, selection, and retention that are central to the rule-following model of decision making. Remove the dependence on predictability from the decision space. Allow decision makers to manipulate the parameters of the decision space. Match the working mental models of business decision makers who look for decisions that fit their external and internal constraints and rules.

GA as a Model of Organizational Change

Their model was similar to other GAs that represent organizational decision making, learning, and evolution. They modeled individual organizations to understand the dynamics of organizational evolution at the population level. Their model represented individual firms as a set of routines. Each allele (e.g., 0, 1, or ?) represented a key strategy routine, and each gene (e.g., 1101?00) represented an organization as a collection of routines.

Bruderer and Singh introduced some variations from earlier models. Specifically, they modeled innovation as a recombination of previous organizational forms, but they did not model organizational aspiration levels. To keep the model simple, they assumed that organizations are always searching to improve.

The model also incorporated intra-generational learning (Baldwin, 1896; Belew, 1990; Hinton & Nowlan, 1987; Morgan, 1896). It is critical to note that this learning introduces a non-Darwinian aspect into the algorithm. The Darwinian explanation of inheritance provides only for translation from the genotype (the genetic structure) to the phenotype (the physical appearance of the individual). This Baldwinian interpretation with intra-generational learning is more in agreement with Lamarck who agreed with Darwin that the genotype affects phenotype. In addition, however, he thought that changes in the phenotype during a life cycle will be represented in the genotype of the next generation. Though this shift is difficult to discern in biological phenomena, it is much more plausible model of decision making. Things learned in one generation will influence the fundamental belief structures and decisions of the next generation.

"Baldwin's (1896) and Morgan's (1896) ideas about how learning guides evolution are crucial, we believe, to the understanding of how a population of organizational forms searches difficult, spike-like fitness landscapes. A key argument is that an evolutionary process is more effective if new organizational forms inherit only some fixed routines from previously successful organizations, while letting others remain open to organizational learning. This allows firms initially to choose some of their key routines incorrectly. Because those routines can be changed later, the organization may eventually discover the correct form. Thus, organizational learning constructs a region of increased fitness around the fitness spike transforming it into a gentler hill. A hill can be searched more effectively by an evolutionary process because the process is constantly guided toward the top of the hill where fitness is greatest . . . . this is like trying to find a needle in a haystack with continual feedback provided about where one is getting closer (Hinton & Nowlan, 1987)." (Bruderer & Singh, 1996)

The modeling approach that captures intra-generational learning includes three kinds of alleles in each gene: 1, 0, and ?. The 1 represents an allele that matches the goal; 0 an allele that does not match the goal; and ? an allele that is undetermined at the time of generation. During the "life" of a single generation, these unspecified traits will be replaced with 1s or 0s in a random search process. This intra-generational determination is the simulation of learning.

Bruderer and Singh (1996) demonstrated that such an algorithm will shorten the time required to find the first fit individual if the landscape is flat with a single fitness peak. Because the GA matches the mental model of rule-following decision processes, includes intra-generational learning and simple inter-generational changes (only crossover), this is the genetic algorithm that is proposed as the basis for a rule-following decision support tool.

GA as a Decision Support Tool

A critical step in the design of any GUI is building the conceptual model. This model identifies the user task objects required to perform the work, all actions, attributes, and relationships among these objects from the users' perspective. The conceptual model also defines a metaphor that captures the task characteristics in an integrated image that is easily comprehended by the user. In later stages of the design and development process, the conceptual design is transformed into a detailed graphical design and implemented as a programmed and functioning application. This paper will deal only with the conceptual model stage of design of the GA decision support tool. This section will outline some of the issues affecting the conceptual design and then outline the conceptual design for the application.

Requirements. Several issues affect the design and will affect the development of this GA decision support tool. These issues will constitute the requirements for the conceptual design of the application. They are described below.

This decision support tool will use the GA as its computational base. Usually, such a decision would not be made until after the conceptual design (based on the users' mental model) was complete. For the purposes of this design the basis of the users' mental model is assumed to match the rule-following models of decision making. Based on the discussion above, the GA is a simulation of this evolutionary process. The underlying assumptions of the GA must be presented in the interface.

Forrester recognizes the issue of scope as a primary concern in building organizational models (Morecroft & Sterman, 1994, 61). The alleles and genes of the GA carry no inherent scale. Each allele could represent a meme, concept, individual, team, department, corporation, industry, nation, or any other size of entity. The audience for the support application will work at many different scales of scope. For example, if a manager is hiring new staff, the alleles might be skills and the genes individual applicants. If a new technology purchase is the question, the alleles might be product features and genes be vendors. If a merger is being considered, the alleles could be corporate departments and the genes the merged entity. Because there is such a wide range of possible applications, the conceptual design must allow for the objects of the algorithm to be assigned a scale by the user at the time of use.

The metaphorical meaning of the gene and its parts are significant aspects of the tool. Bruderer and Singh (1996) defined alleles to be procedures and genes to be corporations. Strict adherence to rule-following model would call them concepts and rules, rules and identities, or identities and decisions (March, 1994). Others who have used genetic algorithms as models of organizational systems have used a variety of other labels and metaphors to infuse meaning into the input and output of the algorithm. In a conceptual design, however, the objects should be drawn from the real-world experience of the users. Most business decision makers do not consciously work at the level of manipulating sets of procedures, nor do they frequently describe their own decision rules and identities. The interface must allow the user to define the objects for manipulation and to provide feedback in the language determined by the user.

The issue of scale also affects time. The GA measures the passage of time by the number of learning trials and the number of generations completed in a given run. These numbers have no absolute meaning for the user in terms of time elapsed. Depending on the decision to be made, one generation may represent 5 seconds or many years. Even though the absolute time is not meaningful to the user, the qualitative proportion of time will be meaningful. If one run of the GA finds a fit in 50 generations and another finds no fit in 1,000 generations, the user can interpret this as a ratio between elapsed times to solution. The tool should allow the user to establish the scale of time for a given run and report data in conformance with that standard.

The tool proposed in this project will not provide all of the possible methods to facilitate rule-following decision-making processes. The rule-following model for decision making provides opportunities for many different decision support tools. Each stage of the decision process might be supported by one or more computer-based tools. Variation, for example might be improved through tools and exercises that spark lateral thinking (de Bono, 1992) or other creative strategies. Selection might include identification of the dimensions that must be fit, comparisons and contrasts on each dimension, methods for making rules and identities explicit, or the ability to delay or defer commitments. Retention would be supported by tools that document past decisions and provide continuous feedback regarding fitness and need for change. The tool proposed here will not focus on optimizing any one step of the process, rather it will allow the user to experiment with parameters that affect the efficiency of the entire evolutionary process.

One of the fundamental differences between the rational choice and the rule-following decision-making models is their definition of a goal. In the rational model, the goal is known by the decision maker. He or she makes the goal explicit as a significant step in the decision-making process. In rule following, on the other hand, the goal is emergent. Decision making involves knowledge of the immediate situation, the patterns of the past, and the capacity of the decision maker. The "goal" is to find the best fit among these components. The goal-as-outcome is unknown until it emerges and proves to be the best fit. For this reason, the GA decision support tool cannot generate for the user a specific outcome solution. Rather, the tool must allow the user to manipulate the conditions under which an outcome will emerge. The tool will help the user experiment with the constraints of the decision space, but it will not generate a solution that will be fit in any given context.

The conceptual model for the GA decision support tool must meet these criteria. The conceptual design is described in the following section.

Conceptual Model. The conceptual model determines the objects, attributes, actions and relationships among the major user task objects and defines a metaphor that integrates all of the objects into a comprehensible whole for the user. Based on the information discussed above, it appears that an appropriate metaphor for the GA decision-making tool will be question and scenario. This concept is used widely in planning and should provide the present- and past-oriented focus that is appropriate in rule-following decision-making processes. A question will represent a real-world issue or decision opportunity that the user wishes to address, and the scenario will represent one set of decision-making circumstances (one run of the GA). When confronted with a decision-making question, the user will generate multiple scenarios and test each to see how quickly a fit is discovered. The user, then, will compare and contrast scenarios to help determine the most appropriate scenario for decision making on the question.

The output for a question would be a sequence of scenarios the user might run. The output from one scenario would be displayed as a graph of the percentages for 0s, 1s, and ?s at each generation of the run displayed on a graph.

This support tool would help decision makers analyze their questions from a rule-following perspective. A decision maker would think about the context and circumstances of the decision and his or her ability to manipulate the decision environment. From this vantage point, the decision maker will be more able to look for a decision that is a best fit for the organization and for persons involved in implementing the decision.

Recommendations for Future Research

Complete high level and detail design, program, and test the decision support tool. The conceptual model builds only the foundation. The design should be tested with potential users to see if the questions make sense to them and if the proposed output is meaningful. When the conceptual design is acceptable, then a high level design (which determines navigation and application look-and-feel) and detail design (which designs the graphical objects) should be completed. Finally the application should be programmed and fully tested for usability and technical and mechanical reliability.

Create a suite of tools that will support each step of the decision process and the process as a whole. This GA tool supports the structuring of the decision space for a particular problem, but it does not influence the effectiveness of the decision maker at each step. A suite of tools might be designed to support all individual steps of the decision-making process. Collect data from decision makers to see how this tool supports individual decisions. In usability tests, laboratory, and working environments, the tool should be tested to determine whether it does conform to the decision-making models people use and whether it improves the efficiency and effectiveness of decisions.

Use it as a teaching tool to help groups manage their own learning and decision making. Because the tool focuses on the decision space, rather than the particular content of the decision, it might be useful as a tool to help decision makers evaluate their own skills and patterns of behavior. Materials might be developed to implement the tool in such a way and research conducted to track improvements in use of the tool and in decision-making behaviors outside of the laboratory.

Such a tool should help distinguish the benefits and limitations of the rule-following decision- making model. Experiments should be designed to test the tool and the model.

Variations can be made in this GA so that it models decision making more accurately. For example, the separate domains that are currently used to determine genesize might be represented instead by separate runs of the GA, each having its own fitness criteria. Or a series of questions or scenarios might be developed from which the user would select the one closest to his or her interests. Many other interesting variations in both the algorithm and the interface would be worth investigation.

Conclusion

Existing decision support tools focus on the rational or limited rationality models of decision making in which the current situation, future goal, and decision maker's preferences are known. This model of decision making is highly reliable, but its validity is questionable in complex, nonlinear, and rapidly changing environments. The rule-following model provides an alternative in which decision makers consider their own identities and their incomplete understandings as they seek to find the best fit between self and environment. This approach to decision making requires a more fluid and emergent tool to support decisions than those based on current decision support technology. The GA, on the other hand, seems to provide the evolutionary foundation for an appropriate decision-making tool to support a rule-following model.

Acknowledgments

References

Anderson, G. & Block, W. (1995). Procrastination, obedience, and public policy: the irrelevance of salience. The American Journal of Economics and Sociology, 54, 201-216.

Bak, P. (1996). How nature works: the science of self-organized criticality. New York: Copernicus.

Baldwin, J. (1896). A new factor in evolution. American Naturalist, 30, 441-451 and 536-553.

Berk, J., E. Hughson, & K. Vandezande. (1996). The price is right, but are the bids? An investigation of rational decision theory. American Economic Review, 86, 954-971.

Bruderer, E. & J. Singh. (1996). Organizational evolution, learning, and selection: a genetic-algorithm-based model. Academy of Management Journal, 39, 1322-1350.

Burnstein, E.; C. Crandall, & S. Kitayama. (1994). Some neo-Darwinian decision rules for altruism: weighing cues for inclusive fitness as a function of the biological importance of the decision. Journal of Personality and Social Psychology, 67, 773-790.

Caster, K. (1994). 15 more management tools for conducting business. PC Magazine, 13, 222-224.

Cohen, M., J. March, & J. Olsen. (1972). A garbage can model of organizational choice. Administrative Science Quarterly, 17, 1-25.

de Bono, E. (1992). Serious creativity: using the power of lateral thinking to create new ideas. New York: HarperBusiness.

Dooley, K. (1997). A complex adaptive systems model of organizational change. Nonlinear Dynamics, Psychology, and Life Sciences, 1, 169-187.

Evans, J. & Over, D. (1996). Rationality in the selection task: epistemic utility versus uncertainty reduction. Psychological Review, 103, 356-364.

Forrester, J. (1994). Policies, decisions, and information sources for modeling. In Morecroft and Sterman (Eds.), Modeling for learning organizations (pp.51-84). Portland, Oregon: Productivity Press.

Fox, C. & Tversky, A. Ambiguity aversion and comparative ignorance. (1995). Quarterly Journal of Economics, 110, 585-604.

Frisch, D. & Clemen, R. (1994). Beyond expected utility: rethinking behavioral decision research. Psychological Bulletin, 116, 46-55.

Gell-Mann, M. (1994). The quark and the jaguar. New York: W. H. Freeman and Company.

Guastello, S. (1995). Chaos, catastrophe, and human affairs: applications of nonlinear dynamics to work, organizations, and social evolution. Mahwah, New Jersey: Lawrence Erlbaum Associates, Publishers.

Hinton, G. & Nowlan, S. (1987). How learning can guide evolution. Complex Systems, 1, 495-502.

Holland, J. (1994). Echoing emergence: objectives, rough definitions, and speculations for ECHO-class models. In Cowan, G., D. Pines, & D. Meltzer (Eds.), Complexity: metaphors, models, and reality (pp. 309-342). New York: Addison-Wesley Publishing Company.

Hollenbeck, J.; Ilgen, D.; Phillips, J.; & Hedlund, J. (1994). Decision risk in dynamic two-stage contexts; beyond the status quo. Journal of Applied Psychology, 79, 592-599.

Kauffman, S. (1995). At home in the universe: the search for the laws of self-organization and complexity. New York: Oxford University Press.

Levinthal, D. (1991). Organizational adaptation and environmental selection--interrelated processes of change. Organization Science, 2, 140-145.

March, J. (1991). Exploration and exploitation in organizational learning. Organization Science, 2, 71-87.

March, J. (1994). A primer on decision making: how decisions happen. New York: The Free Press.

Mellers, B. & Biagini, K. (1994). Similarity and choice. Psychological Review, 101, 505-519.

Morecroft, J. (1983). System dynamics: Portraying bounded rationality. Omega, 11, 131-142.

Morgan, C. (1896). On modification and variation. Science, 4, 733-740.

Roberts, N.; D. Andersen; R. Deal; M. Garet; & W. Shaffer. (1983). Introduction to computer simulation: a system dynamics modeling approach. Portland, Oregon.: Productivity Press.

Simon, H. (1979). Rational decision making in business organizations. American Economic Review, 69, 493-513.

Simon, H. (1982). Models of bounded rationality. Cambridge: MIT Press.

Skyrms, B. Darwin meets "The Logic of Decision": correlation in evolutionary game theory. Philosophy of Science, 61, 503-529.

Slovic, P. (1995). The construction of preference. The American Psychologist, 50, 364-372.

Tversky, A. & Fox, C. (1995). Weighing risk and uncertainty. Psychological Review, 102, 268-283.

Weber, E. (1994). From subjective probabilities to decision weights: the effect of asymmetric loss functions on the evaluation of uncertain outcomes and events. Psychological Bulletin, 115, 228-243.

Wyatt, J. (1995). Nervous about artificial neural networks? The Lancet, 346, 1175-1178.