Multiattribute Decision Making in Context:
A Dynamic Neural Network Methodology

A summary, review, and discussion of concepts described in the article by Leven, S. and Levine, D. (1996) Cognitive Science, Volume 20, Pages 271-299. Abstract: This paper outlines problems that prevent traditional decision making software from functioning well in real world contexts, and presents a theoretical structure for a Multiattribute Decision Making Model.  The theoretical model has a particular focus on incorporating unquantifiable factors (such as psychological) into a Neural Network structure, and Leven & Levine demonstrate the benefits of their theory by relating it to a unique Coca-Cola business scenario. Introduction     The authors Leven & Levine propose a model designed to compensate for those non-quantifiable differences that exist between simulated tests and real world performance. This is referred to as the contextual variance problem: Nonisomorphic performance results whenever a test model fails to accurately represent all essential factors of a 'real world' system, thus fails to be of practical use. To avoid this problem, a model's framework must incorporate those subtle factors that distinguish the controlled laboratory environment from chaotic real-world conditions. Historical attempts to model Stock Market activity provides a good example of contextual variance: Computer programs can accurately predict the value of stock market assets by considering a range of economic variables. And, within a simulated trading environment, it can perform this task flawlessly. The same program, however, cannot predict stock values within the real world market. This is because computer models cannot quantify the human factor: The model cannot compensate for the perceived value of a stock influencing it's actual value. Human activity is difficult to model because factors such as moods, beliefs, and 'gut instinct' speculation elude quantification. And, to further complicate the issue, these influences resonate within the stock market and cause it to rise and fall in unpredictable stochiatic patterns. This global characteristic is not compatible with linear inference engines that use static comparisons against a base line. According to Leven & Levine, this deficiency is what makes traditional programming methods inadequate for complex real world decision making tasks. The Coke Example     A real world business scenario provides an example of how Leven & Levine's model differs from traditional designs. The example is based on Coca Cola's decision in the 1980's to replace traditional Coke with a newly formulated version. According to initial laboratory taste tests, the new Coke outranked the old Coke 2:1; which lead Coca Cola to discontinue the old version and introduce the new version under the same label. Actual consumer response was contrary to the taste test results: The new formulation was so unpopular that Coca Cola had to reinstate the original version as "Coke Classic".     The problem, according to Leven & Levine, was that Coca Cola's market prediction neglected strong psychological factors that were important to the real world consumer. People in the controlled taste test rated new Coke highly; unfortunately, this overwhelming appeal did not exist in the real consumer market because indirect emotional factors, such as memories of expected taste, could influence a consumer's evaluation of the taste. This effect was not only neglected by Coca Cola's testing, but compounded when they ceased to produce the Old Coke and sold the new version under the same label. The Coke label created an expectation of a particular taste, and a familiar sensation, when consumed; and when the product is purchased with this expectation, the new softdrink flavour triggers a negative response to the absence of an expected sensation. This negative psychological response -- frustrative rebound -- is a prime example of nonisomorphic performance of models in real world business decisions. Groundwork for the Model     Leven & Levine outline the standard factors used to uphold contemporary economic theory. They cite Edwards' (1992) Subjective Expected Utility (SEU) theory as the basic heuristics behind modelling traditional economic decision making agents:  1. There exist economic agents;  2. agents have pre-existent preferences among available current and future outcomes;  3. agents independently optimize subject to constraints;  4. choices are made in interrelated markets;  5. agents have full relevant knowledge;  6. observable economic outcomes are coordinated, producing equilibrium states.     The SEU theory is the logic behind economic interactions, however logic alone is not suffient to create successful real world economic decision making agents. This theory does not address a crucial economic factor: Unpredictable human behaviour effects all economic outcomes. This theory needs to include the interplay between external economic variables (such as market context and external inputs) with internal consumer variables (such as beliefs and moods). For an economic model to successfully operate with real world variables, it must consider both psychological and economic market factors. Leven & Levine claim their neural network approach is one way to accomplish this task: "Neural network methods provide not a tight mathematical mapping of effects such as affect and context, but a way to model their significant aspects approximately. Our Coke example illustrates that including such effects can lead to models that fit actual human decision making data better than SEU models do".     To live up to the authors' claim of real world functionality, their model attempts to accommodate three interlinked concepts: Process unquantifiable attributes, minimize nonisomorphic effects, and operate within a dynamic context. The Problem of Dynamic Context     The dynamic nature of real world context is a challenging issue for computer modelling. The authors describe how traditional software designs work well with relatively stable economic factors, but cannot deal with constantly changing aspects. The static approach is relatively simple to design and complies nicely with traditional algorithmic, rule-based, programming techniques; however, in order to operate in a real world context, the model must be able to adapt to chaotic and unstable factors. In this case, the concept of adaption is similar to an optimization process, or making the best of a situation, by determining what goals are realistic and what attributes have an influence on the decision making context. As a contrast to traditional design, Leven & Levine utilize an interactive algorithm concept to enable their model to operate within a dynamic context. They attempt to incorporate a quality where "decisions are based on differential weights of attributes, . . . however, the weights attached to the different attributes can change over time and be influenced by contextual variables". In other words, their model addresses the dynamic context problem by enabling change between multiple goal orientations. Selecting a goal is, in turn, determined by attributes whose proportions reflect the decision making context. Therefore, the neural network's orientation toward a goal can change -- a mild form of adaption -- in order to realize certain goals over others according to context at any point in time. The Problem Nonisomorphic Effects     The authors refer to two technical concepts when they describe the interactive algorithm in their neural network model. The first is neuromodulation, a method of contextual refinement used to compensate for isomorphic changes in attributes (Hestenes, 1992). In the Coke scenario, it represents the change in attribute weighting that occurs as the context shifts from testing to actual buying. In the 'taste test' context the sweetness of taste was emphasized; but in the consumer context, sweetness of taste was secondary to the familiarity of taste. The neuromodulation function is accomplished with an interlevel resonant feedback. This neural network configuration can enable the reality testing of multiple possible goal orientations, thus allowing the system to select the best attribute weighting patterns to accomodate ambiguous values (Hestenes, 1992). For example, it can discern whether New Coke matches the prototype of expected 'Coke-like drinks'; and the degree of match or mismatch can then be used to modify the weights of other attributes (such as taste and familiarity). In this way, the model can establish a weighting scheme from a rigid experimental taste test, and use goal modification to adapt the results to operating with the ambiguous and unpredictable factors relevant to the real world context. The Problem of Unquantifiable Attributes    After outlining the dynamic context issue, Leven & Levine go on to describe how they propose to deal with the difficulty of quantifying psychological attributes. They suggest a way of considering rudimentary emotional influences by using a neural network architecture called the gated dipole. It is important to mention that this design intentionally mimics the neurological principles from psychology's opponent processing theory: "This means that there are representations of pairs of opposites, and that shutting off activity of one of the pair leads to transient activation of its opposite". So, in the Coke example, the authors apply a gated dipole to represent 'security' and 'frustration' as the opponent processing opposites. Upon tasting old Coke, a typical consumer response is familiarity, which is represented by an increase in the weight of the 'security' attribute; which, by virtue of opponent processing, decreases the weight of the 'frustration' attribute. This is an effective technique for avoiding the need to quantify; yet it effectively incorporates the attribute into the model.     From a psychological perspective, Leven & Levine's approach is an attempt to embody the concept of associative learning. Their claim that an interlinked configuration of gated dipoles, across a neural network architecture, can "enable strengthening or weakening of connections between events by contiguity or probable causality" (Hestenes, 1992). In the Coke example, this concept is represented by the learned connections a consumer has with the familiar Coke flavour and a feeling of emotional security. "Just as removal of a negative reinforcer (e.g., electric shock) is positively reinforcing, removal of a positive reinforcer, or its absence when it is expected, is negatively reinforcing". Thus, a person who purchases a Coke has a taste expectation; they associate a particular taste with a feeling of security, and this has been positively reinforced throughout their experience with the product. So, an important factor that motivates a consumer's decision to buy a Coke is their expectation of a feeling of security. Therefore, when such a purchase in made, and the familiar taste is absent, they will not experience the association with emotional security, and instead is left in a negative state of emotional frustration. Discussion     This model is a blend of traditional software architecture and psychological factors; yet it is only a descriptive theory and should be considered as incomplete. There are three areas that are subject to criticism in this model: First, the model serves to outline and collect current ideas, but it does not in itself provide a working representation of such concepts. The model presented is a theoretical logic and only serves to describe how these concepts may be used together in a neural network configuration. Second, the model does not wholly live up to representing psychological concepts, such as associative learning, since it is limited to passive decision making. A creative drive to actively seek out new possible options is a requirement of any system that claims to operate in a truly dynamic context. A third criticism of this model is that it represents the isomorphic Coke scenario in an after-the-fact manner. No evidence is provided about the model's actual predictive ability, however, it remains a useful application of interdisciplinary approaches to understanding intelligence.