Материал: искусственный интеллект

making (also known as preferential choice). These are decisions that do not have an objectively correct answer, but depend instead on the subjective goals and preferences of an individual. Examples include decisions under risk and uncertainty, intertemporal choices, consumer choices, and social interactions. Several new and competing cognitive models have been developed for these tasks, and there is also a rapidly growing body of neuroimaging research supporting these new theories. The purpose of this article is to review these new developments in sequential sampling models for value-based decisions and to stress the importance of identifying cognitive and neural mechanisms that account for the behavioral complexity involved in this type of decision making.

New Developments for Value-based Decisions

Although sequential sampling models for evidenceand value-based tasks share many principles, there are important differences that arise within each type. In addition to the obvious fact that value-based decisions (also known as preferential choices) involve affective rather than inferential evaluations, these tasks usually employ more complex choice options (i.e., options that are characterized by multiple attributes such as the amount and probability of gambles, the amount and delay of intertemporal choice options, or the price and quality of consumer products). These more complex options result in puzzling context effects that seem, from the perspective of economic theory, to be preferential choice paradoxes. Consequently, more elaborate cognitive mechanisms are necessary to account for these paradoxical context effects. Finally, these more elaborate mechanisms are implemented by more intricate networks of neural activity and connectivity profiles.

The application of sequential sampling models to value-based decision making began with decision field theory [13,14], which was initially designed to model binary choices and decision times between uncertain courses of action, and later extended to multi-alternative and multi-attribute tasks (e.g., choosing among consumer products). Shortly afterward, a valuebased version of the leaky competing accumulator model [15], called the multi-alternative leaky competing accumulator model [16,17], was introduced as a competing account for preferential choices. Subsequently, several other competing models appeared, including the attentional drift-diffusion model [18,19], the selective integration model [20], the associative accumulator model [21], the multi-alternative linear ballistic accumulator model [22], and most recently the sequential version of multi-alternative decision by sampling model [23].

Context Effects in Preferential Choice

Why are sequential sampling models needed for preferential choice? One reason is their contribution to understanding how context affects choice. There are many different types of context effects, but, when considering multi-alternative, multi-attribute choices, most research focuses on the effects that ‘choice context’ (defined by the choice set) has on preferences.

Psychologists and consumer behaviorists have long been interested in choice context effects because they provide empirical challenges for economic principles of consistency and coherence that preferences were assumed to obey [24]. These include the principle of independence of irrelevant alternatives (IIA) [25] as well as the regularity principle [26]. According to a probabilistic definition of the independence principle, if A is preferred (chosen more frequently) than B in a binary choice, then A should be preferred (chosen more frequently) than B when a new option C is added to the choice set. According to the regularity principle, the frequency of choosing A in a binary choice between A and B cannot be increased by adding a third option C.

Glossary

Accumulation to threshold models: these models assume that choices are made by accumulating evaluations until a threshold is reached, at which point the choice is made. The same model provides predictions for both choice and decision time. The term ‘sequential sampling models’ is used interchangeably. Approximate Bayesian computation (ABC): a class of computational methods rooted in Bayesian statistics for inferring latent parameters from data by simulating a model to approximate its likelihood function.

Attention switching: refers to the idea that attention switches probabilistically from one attribute to another over time to produce a stochastic sequence of evaluations of different attributes.

Dorsolateral prefrontal cortex (dlPFC): a section of the prefrontal cortex that has been associated with mechanisms of executive control. Recently, it has been linked to nonlinear evidence-accumulation processes in multi-attribute sequential sampling models.

Dorsomedial prefrontal cortex (dmPFC): a section of the prefrontal cortex covering the dorsal part of the cerebral midline that has been associated with cognitive control, conflict, and negative feedback. Recently, it has been linked to evidence accumulation and lateral inhibition processes during decision making.

Independence of irrelevant alternatives (IIA): a principle required by many normative economic choice theories according to which the preference of a person for option A over B should not change if another option C is introduced into the choice set.

Lateral inhibition: the assumption that nodes in a neural network are interconnected by negative links such that positive activation in one node passes negative inhibition to the other nodes. This has the effect of producing a competition that eventually leads to a dominant node of activation.

Loss aversion: this principle defines gains and losses as positive or negative changes with respect to some reference point. The main idea

Studies of choice context provide crucial tests of these normative assumptions. When focusing on choice context effects, most theorists are concerned with three major effects, known as similarity, attraction, and compromise effects (lower panel B of Figure 1, Key Figure). Suppose that option A and option B vary on two attributes (e.g., quality and economics). Option B is economical but low-quality, whereas option A is expensive but high-quality, and, because of the even trade-off, suppose that they are chosen equally frequently.

A similarity effect occurs [27] when a third option, S, which is similar but competitive with option B, is added to the choice set (option S in Figure 1). Adding option S now increases the probability of choosing A relative to B, violating IIA. Interestingly, the similarity effect can be reversed when changing the presentation format from an alternativeto an attribute-focused style [28].

An attraction effect occurs [29–34] when a third option, D, is added to the choice set that is similar but much inferior to option A (option D in Figure 1). Adding option D now increases the probability of choosingoptionAabovethat obtained from thebinarychoice,violatingIIAand theregularityprinciple.

A compromise effect occurs [29,35,36] when a third option C is added to the choice set that extends the attribute values beyond option A, such that option A now appears to be between the attribute values of options C and B, thereby making A appear as a compromise (option C in Figure 1). Adding option C now increases the probability of choosing the compromise option A, violating IIA.

Another type of choice context effect, not shown in Figure 1, but also important, is the reference-point effect [37]. In this case, the decision maker initially is in possession of an option, call it R for the reference, but then must give it up for a new option, either A or B. If R is similar to A, then A is chosen more frequently than B; however, if R is similar to B, then B is chosen more frequently than A. In this case, the location of the reference point reverses preferences between A and B.

Violations of IIA rule out popular models such as the ratio of strengths model [25], and violations of regularity ruled out random utility models [26] as well as the elimination by aspects model [27]. Before the arrival of sequential sampling models of preferential choice, no single traditional static theory could account for all of the above context effects, and sequential sampling models provided the first complete account [38].

Further empirical support for sequential sampling models was obtained by investigating how context effects change as a function of deliberation time [39–41]. As predicted by sequential sampling models, longer deliberation times produce larger attraction and compromise effects (Figure 1, bottom panel). Note that only attraction and compromise effects have been examined using deliberation time manipulations with preferential choice (but see [22] for a study of the effects of deliberation time on similarity effects in a perceptual task).

Boxes 1 and 2 describe some additional findings concerning the three major choice context effects. This includes a summary of substantial individual differences in the different types of context effects (Box 1), and extensions of these context effects to inference tasks (Box 2).

Model Comparisons for Value-Based]FID$T861[ Decisions

Comparison of Model Mechanisms

Figure 1A provides a graphical depiction of a general framework used by the competing sequential sampling models, and Box 3 describes the central mechanisms used in various

is that, when comparing the same magnitude change in gains and losses, the change in loss has a greater impact than the change in gain.

Multi-attribute and multialternative tasks: choice tasks that require participants to choose among more than two options that are characterized by two or more attributes. Context effects are often found in these types of choices. Regularity: adding a new option C to the choice set A + B should not increase the probability to choose either A or B.

Ventromedial prefrontal cortex (vmPFC): a section of the prefrontal cortex covering the ventral part of the cerebral midline that has been associated with the computation of subjective value. It has recently been linked to the accumulation process in value-based decisions.

(A)

30

Threshold

20

Time in steps

A rac on e ect

A

B

D

Compromise e ect

C

B

A

Time steps