The textbook consumer weighs every attribute of every product and picks the option with the highest total value. Research since Herbert Simon’s work on bounded rationality shows that people often do something cheaper. Payne, Bettman, and Johnson (1993) describe an “adaptive decision maker” who holds a repertoire of strategies and switches among them depending on the task, giving up some accuracy to save effort.

Which strategy a person uses matters in practice. A store that knows how its customers decide can offer matching decision support and can predict purchases better. The paper works with four well-studied strategies. Two are compensatory: a weak attribute can be offset by a strong one. Two are non-compensatory: a single attribute can settle the matter regardless of the rest.

Procedural methods watch the process: verbal protocols, mouse-tracking tools such as MouseLab, and eye tracking. They show whether someone searches by attribute or by product and whether the process looks compensatory. They often cannot tell which specific strategy within a category was used, some of them interfere with the decision, and eye tracking is mostly confined to the lab.

Outcome-based methods look only at the final choice and ask which predefined strategy explains it best. Their weakness is that several strategies often predict the same choice. In the ordinary task at the top of this page, a respondent who picks Phone A could be using WAD or MAJ, and the data cannot separate them.

Building tasks in which the candidate strategies predict different choices is called the diagnostic task selection problem (Jekel, Fiedler, & Glöckner, 2011).

The authors want a one-to-one mapping between strategies and alternatives. They call a task’s mapping optimal if each strategy in the predefined set selects exactly one alternative and no two strategies select the same one. Two things can go wrong:

• Ambiguous mapping: two or more strategies select the same alternative. Observing that choice does not identify the strategy.
• Multiple mapping: one strategy is consistent with several alternatives, for example when EBA runs out of attributes with two phones still standing.

Any alternative that no strategy selects is labeled OTHER. A respondent who picks it used none of the four rules. Because utilities, importance weights, and unacceptable levels differ between people, a task that is diagnostic for one respondent is usually not diagnostic for another. Each respondent needs a custom task, built after their preferences have been measured.

The closest earlier method, Euclidean Diagnostic Task Selection (Jekel et al., 2011), enumerates every possible task and keeps the most diagnostic ones. That works for small problems. Their example has two alternatives and four binary attributes, which gives 256 tasks, or 40 after removing symmetric ones. It does not scale to realistic product comparisons.

The procedure has two steps. Step 1 measures each respondent’s preferences: the utility of every attribute level and the importance of every attribute (the authors used adaptive conjoint analysis), plus the levels the respondent finds unacceptable (a questionnaire). Step 2 runs a genetic algorithm (GA) that uses those measurements to search for a choice task with an optimal mapping for that person.

A GA is a heuristic search method. It keeps a population of candidate solutions, favors the better ones, and creates new candidates by recombining and slightly changing existing ones. It does not guarantee the best possible solution, but it can return good solutions for search spaces far too large to enumerate.

Steps 2 to 6 repeat for 600 generations. The encoding is what makes step 6 work: attributes are stored as positions in a list sorted by importance, and levels as positions in a list sorted by utility. A shift of one position therefore produces a task that is slightly different in preference terms, not a random jump.

The mapping score f₁ adds up one term per strategy. A strategy with a multiple mapping contributes nothing. Every other strategy contributes 1 divided by the number of strategies that select the same alternative.

Further goals can be added as extra terms. In the study the objective was f = f₁/2 + f₂ + f₃, where f₂ and f₃ (each between 0 and 1) reward hitting a target level of complexity, explained in the next section.

Robustness. Measured preferences contain error, so a mapping that breaks when one value shifts is of little use. After the last generation, each of the 300 archived tasks is copied 50 times with one randomly changed attribute level, and the objective is averaged over the copies. The task with the highest average, written f′, is the one shown to the respondent.

Most studies of choice difficulty vary the size of the task: more products, more attributes. The paper instead manipulates context-based complexity, which depends on how the attribute levels relate to each other in the eyes of the respondent. Two measures are used.

Similarity is measured by the attribute range (AR): the standard deviation of utilities across alternatives, averaged over attributes. A low AR means the products are close in value on every attribute, so none can be ruled out quickly.

Conflict is measured by the average pairwise correlation between attributes (AC). A negative correlation means a product that is good on one attribute tends to be poor on another, so choosing requires trade-offs. A positive correlation means good things come together and some product tends to dominate.

Earlier process-tracing studies suggested that harder tasks lead to more thorough, compensatory processing. Outcome-based studies suggested the opposite. The authors side with the second view, following Bettman et al. (1993): people may aim for accuracy but cannot carry out the required calculations, and fall back on simpler rules.

This demo implements the algorithm as described in the paper and its appendices, on the 15 phone attributes of the study. The respondent is synthetic: importance weights, utilities, and unacceptable levels are drawn at random, because the study’s individual data are not public. Random sampling with the same number of evaluated tasks runs alongside as the benchmark, as in the paper.

The authors measured preferences for 301 people in an online study in Japan in 2010 (all aged 30 to 39, all phone owners, 60.5% women) and generated tasks with five or eight alternatives and five or eight attributes. The benchmark drew 180,000 random tasks per design, the same number the GA evaluates.

With complexity targets, about 10% of the GA’s tasks did not have an optimal mapping. The authors give two reasons: for some preference profiles no such task may exist, and the search sometimes gives up mapping quality to hit the complexity targets. These tasks were still used. A choice is informative whenever the chosen alternative is selected by exactly one strategy, even if other alternatives in the task are ambiguous.

Generating the 4,816 tasks of the main experiment took 1.21 minutes per task on average, or 27.09 hours in total on four parallel processors.

Each of the 301 respondents made 16 choices: every combination of low or high similarity, low or high conflict, five or eight alternatives, and five or eight attributes. Average decision time was 23.4 seconds. Choices that could be explained by more than one strategy, or by a strategy with a multiple mapping, were removed, leaving 4,567 of 4,816 observations. The analysis is a multinomial multilevel regression (GLLAMM in Stata) with a random effect per respondent.

• Context-based complexity mattered more than task size. Adding the number of attributes did not improve model fit (Wald test, p = .21). The number of alternatives did, but only through OTHER, whose odds rose by about 50%. That is largely mechanical: with eight alternatives and an optimal mapping, four alternatives count as OTHER instead of one.
• OTHER grew with both kinds of complexity. Its odds relative to WAD roughly doubled under high similarity (2.12) and under high conflict (2.25).
• Interaction. In the model with a similarity × conflict term, the effect of conflict on MAJ is stronger when similarity is low (interaction odds ratio 1.38, p = .096, which is above the conventional .05 threshold).
• Overall pattern. Conflict mainly moves people toward EBA and away from MAJ. Similarity mainly moves people away from WAD. The authors see this as consistent with Bettman et al. (1993): as complexity rises, people are less able to carry out demanding compensatory strategies.

• One alternative per strategy. An unambiguous mapping needs at least as many alternatives as strategies. Two proposed ways around this: add process measures such as eye tracking so that overlapping strategies can be separated by process, or let the GA design a portfolio of tasks per condition.
• Pure strategies only. People may combine strategies, for example screening with EBA and then finishing with WAD. The design does not detect mixtures. The large OTHER share in the hardest tasks is consistent with mixing.
• LEX was barely above chance. The authors plan to replace it and to test more strategies.
• Preferences are assumed to be stable. The tasks depend on preferences measured beforehand. If preferences are constructed during the choice, as much of the literature holds, the mapping can be off. The robustness step addresses small measurement errors only.
• No guarantee of optimality. The GA is a heuristic. In the complexity experiment its robustness score f′ was about the same as that of random sampling (2.37 versus 2.36).
• One sample, one product category. The data come from 301 Japanese respondents aged 30 to 39 choosing among cell phones in 2010.