Материал: искусственный интеллект

finally met by advancing value-based sequential sampling models that achieve this capability by introducing a variety of new mechanisms (distance-dependent lateral inhibition, loss aversion, attention guided by rank value). Almost all the value-based sequential sampling models can account for similarity, attraction, compromise, and reference-point effects (except for [19] and [20]). In addition, decision field theory (based on lateral inhibition) and the leaky accumulator model (based on loss aversion) provide strong a priori reasons for the observed negative correlation between attraction/compromise effects and similarity effects (Box 1). The associative accumulation model provides the most systematic account of reference-point effects [43] so far. However, the decision by sampling model [23], being the most recent application to context effects, accounts for the largest number of different qualitative findings (see the 25 different phenomena listed in Table 4 of their article))]FID$T17[.

A crucial dynamic prediction made by the sequential sampling models concerns the temporal evolution of preferences. The sequential sampling models also predict that attraction and compromise effects grow larger as a function of increasing deliberation time (Figure 1B), and the predicted increasing effect of deliberation time has been confirmed in several experiments [39–41]. The dynamic nature of these effects is important because several new static choice models of context effects have been proposed [36,44–47] that have no mechanisms for making any a priori predictions about dynamic effects.

Quantitative Empirical Comparisons

Several quantitative model comparisons have been conducted to compare the accuracy of the competing sequential sampling models for predicting context effects in value-based choice. The models have been compared using several different methods (Box 4). Some comparisons are based on using aggregate data (pooled across participants), others are based on predictions for individual data, and finally some use hierarchical methods that apply to all participants by including an additional model for the distribution of individual differences. Table 2[172T$DIF]provides a summary of the model comparisons. Note that this only includes comparisons based on preferential choices among value-based options, and does not include comparisons based on perceptual or inference tasks ([22] gives an example of the latter). Also note that, although all the sequential sampling models are capable of predicting both choice probability and decision time, the comparisons shown in Table 2[173T$DIF] are based only on choice data.

Box 4. Methods for Evaluating Models

Quantitatively evaluating the predictions of cognitive models for empirical data usually requires estimating model parameters from part of the data. Bayesian estimation methods have become more popular in cognitive science because they enable hierarchical versions of cognitive models to be fit to the data, and many methods and software packages have facilitated this transition [96]. Once fit, researchers can use methods to obtain metrics such as Bayes factors [97] to assess the evidence for one model or another. However, the complexities of the models described in this article make it difficult to derive simple equations for model fitting, making parameters difficult to estimate. This situation is problematic because it prohibits researchers from using parameter estimates to characterize individual differences, and understand what combination of model mechanisms yields specific patterns of behavioral data. Fortunately, new methods of parameter estimation circumvent the complex mathematical details of the models through model simulation [98]. Often referred to as approximate Bayesian computation (ABC), these methods take summary statistics of simulated data, compare them to observed data, and use the discrepancy between the two statistics as a measure of how likely each model parameter is to have generated the observed data. The novelty of the ABC approach is that it can be used within a Bayesian framework, and thus hierarchical models and parameter uncertainty can easily be assessed. Many new algorithms have been developed for specific modeling applications, such as estimating parameters that are intercorrelated [96,99], models of choice response time [100,101], recognition memory [102], preferential choice [48], and hierarchical models [103]. Together, these algorithms have opened up new opportunities for assessing complex individual differences, as well as comparing model fit, balanced for model complexity.

Trends in Cognitive Sciences, March 2019, Vol. 23, No. 3 257

Table 2. Comparison of Competing Models with Respect to the Accuracy of Quantitative Predictions for]FID$T61[ Preferential Choicesa,b

aAbbreviations: AA, associative accumulation model [21]; DbS, decision by sampling model [23]; DFT, decision field theory [38]; LBA, linear ballistic accumulator model [22]; LCA, leaky competing accumulator model [16]; MNL, multinomial logit model [26].

bNote that the attentional drift-diffusion model [19] and the selective integration model are not included because they have not been quantitatively compared with other models with respect to predictions for the main three choice context effects.

The results of the competition show that sequential sampling models generally make better predictions than the multinomial logit model (a popular random utility model). However, the results of competition among sequential sampling models indicates that, although decision field theory performs well for aggregate data, other competing models perform better when individual differences are taken into account.

A better way to evaluate the sequential sampling models is to form a new collection by systematically including or excluding the various component processes that are used in different existing models (Box 3). Using this strategy, it is possible to identify the crucial psychological mechanisms important to the decision, rather than any specific ensemble of mechanisms assumed by a particular model. Recently, Turner and colleagues [48] compared a collection formed by including or excluding different types of attention shifting/weighting, loss aversion, lateral inhibition, and noise assumptions. The results of this large ‘switchboard analysis’ of model comparisons indicated that the best-performing models include stochastic integration of attribute comparisons, attention weighting depending on attribute values, lateral inhibition, and non-linear evaluation of attribute comparisons.

It seems difficult to distinguish the sequential sampling models on the basis of choice data alone. However, an added advantage of these models is that they also predict decision time and derive implications for eye movements, which can also be used for model comparison. There are numerous applications of value-based sequential sampling models to choice and decision time for the simple case of binary choices []FID$T471[13,18,19,49–52], but fewer applications multi-alternative (more the two) choices []FID$T571[53]. By adding additional assumptions linking eye movements to attention, predictions can made regarding the direction of eye movements during the decision process [54] and the influence that this direction has on choice [18,19,41]. Another important way to distinguish between models may be obtained from neuroscientific evidence [55], as reviewed in the following section.

Neuroscientific Research on Mechanisms

Neural Mechanisms of Value Accumulation

Early neuroscientific studies that relied on sequential sampling models to better understand value-based decisions focused on the question of which brain regions mediate the processes of value integration and evidence accumulation [56–59]. Consistent with other work in

258 Trends in Cognitive Sciences, March 2019, Vol. 23, No. 3

neuroeconomics [60,61], these studies agreed on the role of the ventromedial prefrontal cortex (vmPFC) as representing the subjective value of available choice options (middle panel of Figure 1B). However, although some studies further linked the vmPFC to comparison and evidence-accumulation processes [59,62], other studies attributed these cognitive mechanisms to downstream areas such as the dorsolateral prefrontal cortex (dlPFC) and dorsomedial prefrontal cortex (dmPFC) (right panel of Figure 1B) [57,58]. Importantly, these early as well as many more recent studies in decision neuroscience [63–66] assume that people accumulate and compare integrated value signals of each option, which stands in contrast to the idea of stochastic switching across attributes, a mechanism inherent to most theories of multi-attribute decision making (see above). One reason for this discrepancy appears to be the dominance of fMRI as a tool to study the neural basis of value-based decision making in humans. The low temporal resolution of fMRI does not allow us to measure rapid changes in attention and decision processes. In our view, future research will need to rely more heavily on techniques with higher temporal precision such as electroencephalography (EEG) and magnetoencephalography (MEG) to study value-based decisions. Furthermore, the integration of knowledge from animal studies that employ rapid single-unit recording as well as optogenetic interventions will be crucial [10,67,68] even though this research area has mostly focused on perceptual decisions so far. Another reason for the discrepancy between cognitive and neural studies of value-based choice seems to be the frequent use of choice stimuli with ambiguous attributes (e.g., food snacks) in decision neuroscience, which precludes measuring and dissociating attribute-specific computations.

Neuroimaging Studies of Context Effects and Attribute-Wise Decision Processes

Despite the modest take-up of insights from the cognitive sciences, a few neuroscientific studies have investigated context effects in multi-attribute decisions, and especially the attraction effect [69–73]. Two studies reported increased activation of the anterior insula, either when contrasting target against competitor choices [71] or when contrasting decisions with a decoy option against decisions with a neutral third option [72]. The involvement of the anterior insula may indicate that saliency-driven overweighting of the strongest attribute of the target underlies the attraction effect [21,72,74].

Only a subset of these studies made use of value-based sequential sampling models to connect the neural data with potential cognitive mechanisms that underlie the contexts effects in multi-attribute, multi-alternative choice tasks [72,73,75]. One of the first was an fMRI study investigating the neural mechanisms of changes in attribute relevance in a threealternative choice task [75]. In this study, an attribute became more relevant (i.e., had a stronger influence on the decision) if one of the options had an exceedingly high value on this attribute. To explain the (IIA-violating) choice behavior in their task, the authors used a hierarchical accumulator model that bears many similarities to multi-alternative decision field theory [38], although it would not be sufficient to explain all the above-mentioned context effects (i.e., attraction, similarity, compromise). At the neural level, it was found that the ventromedial prefrontal cortex and the intraparietal sulcus encoded a chosen-value signal that was modulated by attribute relevance, while the dorsomedial prefrontal cortex encoded an unmodulated value signal. A study more directly concerned with the attraction effect [72] applied multi-alternative decision field theory [38] to predict choices between risky prospects. The predicted choice probability of the model could be linked to fMRI activation in vmPFC and posterior cingulate cortex (PCC). Interestingly, the choice-related activity in PCC was stronger in those participants who – according to the cognitive model – exaggerated the psychological distance between the target and the decoy in the 2D attribute space.

Trends in Cognitive Sciences, March 2019, Vol. 23, No. 3 259

260 Trends in Cognitive Sciences, March 2019, Vol. 23, No. 3

Finally, some initial progress has been made recently in uncovering the neural substrates underlying the advanced mechanisms used by sequential sampling models of value-based decision, for instance by linking lateral inhibition of choice options to the cognitive control system of the brain [76]. However, most of the advanced mechanisms that have been put forward in cognitive research on value-based decisions making are yet to be linked to brain data. More work and effort is needed in this regard, and we have delineated an agenda for future research. The ultimate goal is use neural data to constrain cognitive models [83] to identify a neurobiologically plausible account of value-based decision making and to avoid further inflation of proposals of sequential sampling models that are able to explain context effects.

Acknowledgments

J.R.B. was supported by the Air Force Office of Scientific Research (FA9550-15-1-0343), S.G, and J.R. were supported by

the Swiss National Science Foundation (100014-172761 and 100014-153616, respectively)]FID$T71[.

Trends in Cognitive Sciences, March 2019, Vol. 23, No. 3 261