Why does the tendency to define problems in terms of a preferred solution occur?

Human decision making is often studied as the outcome of a careful evaluation of alternative options in terms of the likelihood and the value of outcomes associated with these options. Subjective Expected Utility theory is probably the most widely used normative theory that is based on these assumptions. A considerable amount of research focuses on systematic violations of these principles of rationality in order to help understand the cognitive processes that underlie human judgment and decision making. Research on heuristics refers to cognitive short cuts that people use in their decision making. The most important heuristics are briefly described, followed by information about more descriptive approaches to decision making (e.g., prospect theory, image theory, rule following, reason-based choice). Finally, the possible contribution of formulas or simple linear models to the quality of human decision making is discussed briefly.

Important Takeaways

•

Human decision-making is strongly biased by unconscious mental processes (system one) that sometimes produce good outcomes quickly but sometimes cause us to make irrational choices. Our rational mind (system two) rarely intervenes.

Fear of loss influences human decisions more than expectation of gains. This bias affects people’s choices in risky situations, like whether to buy insurance, accept lawsuit settlements, gamble, or skydive. Psychologists Kahneman and Tversky conducted experiments demonstrating the pervasiveness and strength of this bias.

Framing—how a choice is worded—affects how people choose. People prefer a sure thing over a gamble when options are worded as gains, and gambles over sure things when the identical options are worded as losses. This bias makes people susceptible to anchoring: a mind trick where someone sets your expectations to a certain level, then shows you either how to improve your outcome or how to avoid a worse outcome.

People are biased toward options that are easier to recall or envision. A close relative’s experience with a product influences our willingness to buy it much more than reading statistics or online reviews about the product.

Our past decisions bias our future ones, because people try to behave consistently. Therefore, people tend to stay with what is familiar, stick with losing causes longer than they should, and like things better if they put more effort into getting them.

Emotions are critical to decision-making. Without an emotional response of some sort, it is difficult to make decisions.

Designing to exploit strengths and weaknesses of human decision-making:

Support rational decision-making: Help system two override or co-opt system one by providing all options, showing alternatives, providing unbiased data, performing calculations for users rather than forcing them to calculate, and checking the assumptions underlying the reasoning.

Make AI-based systems more transparent.

Use data visualization to harness system one to support system two.

Use persuasion ethically. Don’t influence people to do what is contrary to their own interests.

14.4.1 Overview

Research in human decision-making has revealed the fact that people have a strong tendency to keep the status quo when choosing among alternatives (see, e.g., Kahneman et al., 1991; Samuelson and Zeckhauser, 1988). This effect is known as status quo bias. A consequence of this is that proposed decisions (e.g., decisions proposed by experts or by the configurator application) that represent the status quo are accepted by the user. Defaults (see, e.g., Tiihonen et al., 20145) can lead to such a status quo bias. The results of the research of Samuelson and Zeckhauser (1988) implied that an alternative was significantly more often chosen when it was designated as the status quo, and that the status quo effect increases with the number of alternatives. Kahneman et al. (1991) argue that the status quo bias can be explained by a notion of loss aversion, since the status quo serves as a neutral reference point, and users evaluate options in terms of gains and losses relative to the reference point. Since individuals tend to regard losses as more important than gains in decision-making under risk (i.e., alternatives with uncertain outcomes; Kahneman and Tversky, 1979) the possible disadvantages outweigh the advantages. Cosley et al. (2003) showed that presenting item ratings in collaborative filtering recommenders (Konstan et al., 1997) has an impact on the rating behavior of a user. For example, ratings were higher in situations where inflated predictions were presented to the user.

11.5.1 Norm Assistance

In certain scenarios, human decision making is affected by policies, often represented by deontic concepts such as permissions, obligations, and prohibitions. Individual rules within a policy have been actively studied in the area of normative reasoning [14]. Norms generally define constraints that should be followed by the members in a society at particular points in time to ensure certain systemwide properties [12]. These constraints are generally specified by guarded logic rules of the form ν←ϕ, which indicate that when the condition ϕ occurs, a norm ν becomes activated, imposing a restriction on the set of desirable states in the domain. If ν is an obligation, the norm defines a set of states through which an agent must pass; otherwise, if ν is a prohibition, ν the norm defines a set of states that must be avoided.

For example, in international peacekeeping operations, military planners must achieve their own unit’s objectives while following standing policies that regulate how interaction and collaboration with non-governmental organizations (NGOs) ought to take place. Because the planners are cognitively overloaded with mission-specific objectives, such normative stipulations increase the complexity of planning to both accomplish goals and abide by the norms.

Although much of the research on normative reasoning focuses on deterministic environments populated by predictable agent decision making, such a model is not suitable for reasoning about human agents acting in the real world. By leveraging recent work on normative reasoning over MDPs [9], it is possible to reason about norm compliance in nondeterministic environments; however, the issue of nondeterminism in the decision maker has remained problematic. To address this problem, an instantiation of the proactive assistance architecture was created to provide prognostic reasoning support by designing the proactive manager to analyze user plans for normative violations [21] in the context of military escort requests for relief operations. An overview of this architecture is provided in Figure 11.2a, while a screenshot of the assistance application is shown in Figure 11.2b.

The normative assistant relies on a probabilistic plan recognizer to generate a tree of possible plan steps. The proactive manager evaluates a user plan through a norm reasoner, which analyzes the sequence of states induced by the predicted plan for norm violations. These predicted violations are the object of the agent planner, which tries to find the nearest norm-compliant states in order to recommend user actions that will ensure norm-compliant behavior. If compliant states are not achievable, for example, because some violations are unavoidable in the user’s possible future state, or if the user has already violated certain norms, the agent can also suggest remedial actions to either comply with penalties or mitigate the effects of a violation (i.e., contrary-to-duty obligations [23]).

6.3.2 BI as a Decision Support Tool

Our primary focus up to this point has been on BI as a strategically important capability that is leveraged to enhance the efficiency and effectiveness of critical business processes. While we have mentioned that BI is a decision support tool used within those business processes, we have not explicitly looked at how BI works in the context of business decision-making. Since business people make all kinds of decisions every day, and since our focus is on business performance and business processes, we’ll narrow our focus to the use of BI to support business decisions about how best to respond to performance issues. As we discussed early on, BI is an umbrella term that encompasses several different styles of BI. These range from static reports to highly sophisticated analytics, and their uses for decision support vary—as shown by Table 6.7.

Table 6.7. BI Can Be Used Across the Stages of Business Decision-Making

While there are many models of human decision-making, Table 6.7 reflects a simple framework of decision stages—a framework that will be useful for showing how BI can be used to support decisions for resolving business performance issues. The model assumes the following decision stages:

Recognition that there is a performance issue that needs to be resolved—whether through a regular periodic business performance review process or through ongoing performance monitoring;

Determination of the Importance of the performance issue, whether that is characterized in terms of financial impact, customer service impact, or other relevant performance measure—so as to focus managerial bandwidth and attention appropriately;

Developing an understanding of Causal Factors of the performance issues;

Formulation and evaluation of Options; and

Putting forth a Recommendation.

For each of the decision stages, we see that various styles of BI can be useful. The old standby for most companies is the static report, which has utility but does not capitalize on modern BI-enabled decision support capabilities. Most executives and managers want to understand what happened in the past, but they are far more interested in being able to predict the future, assess their options, and deal with the complexities of the decisions to be made. Leveraging the right BI tools provides the kinds of decision support that executives, managers, and analysts say they need to make more impactful and timely decisions.

Model-Based Interactive Planning

For an increasing number of applications human decision-making is supported by a computer simulation of a system. The model forecasts the behavior of a system by starting with a set of initial conditions and running it forward in time for a period which may range from hours to many years. The user operates in an interactive loop, adjusting model parameters and examining the predicted outcome. In this way the flexibility and subtlety of human understanding can be combined with computation power in a distributed cognitive system.

Example 1. A business spreadsheet is used to forecast profit margins based on assumptions about such variables as materials costs, labor costs, cost of transportation, etc. The output is a time series of forecast revenues and profits.

Example 2. A traffic behavior model is used to simulate the effect of alternative intersection designs on congestion patterns.

Example 3. A model of interactions between a set of commercial fish species is used to forecast the effects of changing the allowable catch.

Fig. 12.15 shows an interface to the MSPROD fisheries model (St Jean, Ware, & Gamble, 2016). The analyst can change the fishing effort on different species categories and immediately see the resulting change in a 30-year forecast of the fish populations. The shaded portion of each time series plot shows the change from a previous forecast. The model is based on known predation of one species on another, for example, dogfish each lots of cod, as well as on competition between species; for example, haddock and cod are both groundfish competing for the same underlying food resources. A user such as a fisheries administrator can change the amount of allowable catch on a particular fish species and see the forecast effects of this change for all ten species.

One of the problems with computer-based modeling is that most systems only show results and the inner workings of the model are invisible. The system shown in Fig. 12.15 is an attempt to counter this problem by showing the chain of cause and effect from one species to another. The arcs to the left and right show interactions between fish species representing competition and predation. An increase in the catch of elasmobranchs (dogfish and skates) results in a decrease in the populations of those species and a considerable increase in cod and winter flounder. The latter benefit because they are a favorite food of elasmobranchs. But windowpane stocks decrease because they compete with cod. The dynamically sized links reveal the inner workings of the model and allow for an analyst to not only see that a forecast has changed, but also to understand the reason for the change.

Design Guideline. As far as possible, make the underlying workings of the model visible to the user so that results can be explained.

In many cases, computer models of systems include uncertainties in the predictions. The problem of displaying these uncertainties has already been discussed in Chapter 9 for hurricane simulations. In generally it is always important to display model uncertainties, especially in the case the area large. The system shown in Fig. 12.13 had an alternative view in which uncertainties where shown—in the case of the fisheries model, the uncertainties were very large.

Cognitive Guideline. Make model uncertainties visible, so that decision makers can take these into account.

9.9.4 Appendix 4: Selten’s view of bounded rationality

Modern mainstream economic theory is largely based on an unrealistic picture of human decision making (Selten, 1998, 1999). Economic agents are portrayed as fully rational Bayesian maximizers of subjective utility. This view of economics is not based on empirical evidence, but rather on simultaneous axiomization of utility and subjective probability. However, following Savage, Selten states that “the axioms are consistency requirements on actions with actions defined as mappings from states of the world to consequences” (Savage, 1954). In Selten’s opinion, the imposing structure built by Savage has a strong intellectual appeal as a concept of ideal rationality. However, it is incorrect to assume that human beings conform to this ideal. The split of the person into multiple selves with conflicting goals in itself is a bound of rationality for the person as a whole, even if it is not cognitive but motivational. Not only cognitive, but also motivational bounds of rationality must be taken into account by a comprehensive theory of bounded rationality (Selten, 1999).

The further reasoning of Selten adduces Becker (1967) and Becker and Leopold (1996) who have proposed a theory of household behavior which extends aspiration adaptation theory to the right context. The household divides its monthly income into a number of funds for different kinds of expenditures like a fund for food, a fund for clothing, a fund for entertainment, etc. The goal variables are the fund sizes and upper price limits for wants, like the desire for a pair of shoes or an excursion advertised by a travel agency. Such wants are produced by a want generator, modeled as a random mechanism. When a want is generated by the want generator, another instance, the administrator, checks whether there is still enough money in the appropriate fund and whether the want remains under the price limit for such desires.

If the price limit is violated, the want is rejected. If the want remains under the price limit but there is not enough money in the fund, then the want will still be granted if transfer rules permit the transfer of the missing amount from another fund. (The structure of these transfer rules is not explained.) If such a transfer is not permissible, then the want is rejected. At the end of the spending period, a new aspiration level for the next one is formed by aspiration adaptation in the light of recent experience. If the household theory of Becker (1967) is applied to the spending behavior of a single person, then want generator and administrator are different personality components. Conflicts between them are not modeled by the theory, but it may be possible to extend the theory in this direction. Everyday experience suggests that sometimes wants are realized against the will of the administrator. The split of a person into a mechanistically responding want generator and a boundedly rational administrator seems to be a promising modeling approach not only to household theory, but also for other areas of decision making.

In conclusion, Selten (1999) hopes that he succeeded in conveying the essential features of bounded rationality. In the introductory part, he argued that rational decision making within the cognitive bounds of human beings must be nonoptimizing. The exposition of aspiration adaptation theory served the purpose of demonstrating the possibility of a coherent modeling approach to nonoptimizing but, nevertheless, systematic and reasonable boundedly rational behavior.

Finally, Selten (1999) writes: “What is bounded rationality? A complete answer to this question cannot be given in the present state of the art. However, empirical findings put limits to the concept and they indicate in which direction further inquiry should go.” For other approaches and opinions see, in particular, Pattillo (1977).

The issue of bounded rationality is still vital. Read current books and the most recent papers on newest writings pro- and contra-bounded rationality after the publication date of the present book.

Fuzzy Inference Engine

The fuzzy inference engine employs a particular kind of fuzzy logic. It simulates human decision-making procedures, and employs a fuzzy knowledge-base and fuzzy input to generate fuzzy decisions (output).

There are two common methods for performing fuzzy logic inferences: the max-min method and the max-product method. In the max-min method the final output membership function for each output is the union of the fuzzy sets assigned to that output, and the degree-of-membership values are clipped at the degree-of-membership for the corresponding premise. In the max-product inference method, the final output membership function for each output is the union of the fuzzy sets assigned to that output in a conclusion, and their degree-of-membership values are scaled to peak at the degree-of-membership for the corresponding premise [7].

3.1 Framing

In a study that is widely believed to illustrate a deeply irrational feature of human decision-making, Tversky and Kahneman (1981) presented a group of subjects with the following problem:

Imagine that the US is preparing for the outbreak of an unusual Asian disease, which is expected to kill 600 people. Two alternative programs to combat the disease have been proposed. Assume that the exact scientific estimate of the consequences of the programs are as follows:

If Program A is adopted, 200 people will be saved.

If Program B is adopted, there is a 1/3 probability that 600 people will be saved, and a 2/3 probability that no people will be saved.

A second group of subjects was given an identical problem, except that the programs were described as follows:

If Program C is adopted, 400 people will die.

If Program D is adopted, there is a 1/3 probability that nobody will die and a 2/3 probability that 600 people will die.

On the first version of the problem most subjects chose Program A. But on the second version most chose Program D, despite the fact that the outcome described in A is identical to the one described in C.

8.4 Synergetic links

The behavioral models that we present in this chapter comprise stochastic operational models that mimic human decision-making to predict human behavior and fulfill the quantitative requirements, such as low-latency and energy consumption, a TI application has to meet. Such a model is of utmost importance in many TaHiL use cases: For robot-assisted medical and industrial training issued in the research groups for medical applications (U1, see Chapter 2), industrial applications (U2, see Chapter 3), as well as in the Internet of Skills (U3, see Chapter 4).

In robot-assisted environments (U1 and U2), the key requirements that the model needs to fulfill are efficiency and suitable explanations to support the human with key information about which tasks will be achieved by the robotic assistant and what is left to the surgeon or industrial worker. To this end, our models might provide the basis to decide when to handover control from machines to humans, and vice versa. Also in critical situations where coordination of all participants (human and robotic) is important, our models could provide predictions on how humans could react and how complications could be avoided.

The human-to-machine transfer via modeling illustrated in Chapter 2 could be also supported by learning human-inspired models. Here, context-aware assistance is in the line of selecting contexts (see Section 8.2.7) and policies of the learned library component (see Section 8.3.3.2).

The Internet of Skills (U3, see Chapter 4) further relies on CPSs to be developed. Following the context-adaptive software concepts (TP5, see Section 13.5), the development of such systems will rely on a model-based approach such that also operational models are at hand. These models could be included into our learning framework to analyze policies depending on the context of the CPS. For using machines to support learning in the Internet of Skills, it is important that machines adapt themselves to the predicted behavior and demands of an individual user. In these applications, our human-inspired learning approach could continuously learn the behavior of the humans and reason about their selected policies and habits.

Besides the general human-inspired learning approach, additional knowledge and refined versions of the models could be included to instantiate general human-inspired models to specific individuals. This addresses also the instantiation of models for the whole lifespan of humans, as illustrated in Chapter 9. For instance, elderly people might stick more to their previously learned strategies, whereas young people may tend to experiment more. This is important not only in rehabilitation use cases (U1, see Chapter 2), but also for teaching scenarios in the Internet of Skills (U3, see Chapter 4). To instantiate learned models for an individual Human-in-the-Loop, classical learning techniques [540,541] can be employed to extract key features of the individual. The long-term goal is to enable automatic system adaptations as well as providing learning and explanatory facilities that fit individual demands and objectives of the user. This goal requires joint forces, e.g., relying on essential technologies for tactile computing issued in Chapter 13. Furthermore, results from TP1 and TP2 (see Chapters 9 and 10) could help to provide input for training our human-inspired reasoning models.