Strategy use is influenced by various dispositional and situational factors. Early studies identified that learning related factors—expertise, training, and prior knowledge of the decision problem—impact predecisional information search and strategy use (Zakay, 1990; Shanteau, 1992; for recent work see Pachur and Marinello, 2013). Later, interest grew in the relation between strategy use and motivational characteristics such as need for cognitive closure (Shiloh et al., 2001; Vermeir et al., 2002), need for cognition (Verplanken, 1993), decisiveness and indecisiveness (Ferrari and Dovidio, 2001; Anderson, 2003). Also, cognitive ability was found to impact strategy selection (Bröder, 2003; Fasolo et al., 2003; Mata et al., 2007, 2010). Numerous task and context determinants of predecisional information processing and strategy use studied so far include framing of a decision task (Tversky and Kahneman, 1981), response mode (Tversky et al., 1988), type of learning task (Pachur and Olsson, 2012), information cost and information format (Bröder, 2000; Newell and Shanks, 2003); task complexity (Payne, 1976), time pressure (Wright, 1974; Rieskamp and Hoffrage, 1999), and many others (for overviews see Payne et al., 1993; Gigerenzer and Gaissmaier, 2011).

Notably, very few studies looked at how emotional factors influence predecisional information processing and strategy use. Mano (1992) reported that participants under social stress were more extreme in their judgments of job applicants, suggesting that they may have focused on only a subset of the presented information. Similarly, Lewinsohn and Mano (1993) showed that highly aroused participants acquired less information and did so in a more selective manner by focusing more on subjectively important attributes. Scheibehenne and von Helversen (2015) looked at how positive and negative mood states impact strategy selection, assuming that positive mood broadens the attention focus and negative mood narrows it. In line with this assumption they found that positive mood led to the more frequent use of compensatory strategies, whereas negative mood led to the more frequent use of simple noncompensatory heuristics. Wichary et al. (2015a) looked at how emotional arousal impacts the use of decision strategies. In that study, we used skin conductance as an index of autonomic arousal induced by the aversive and neutral pictures, in order to better control participants' state and to gain some insight into the neural correlates of predecisional information processing. We found that our manipulation led to reduced information search and increased allocation of attention to the most important cue. We also found that emotional arousal indexed by skin conductance leads to the use of simpler decision strategies. This pattern of results favors the attention narrowing hypothesis, which suggests that high arousal can lead to attention narrowing and in consequence, to the reliance on simple decision heuristics. The association of strategy use with skin conductance is also a hint at the neural underpinnings of strategy use, both peripheral and central. Our results indicate that the sympatho-adrenergic-medullary system (SAM; Schommer et al., 2008) is involved in decision processes under stress, as well as that the central structures associated with SAM activity might be involved, namely the anterior cingulate cortex (Critchley, 2005) and the neuromodulatory noradrenergic system (Aston-Jones and Cohen, 2005; Nieuwenhuis et al., 2011).

Insights into the neural mechanisms of strategy selection come from a separate line of research within neuroscience. Studies on multi-attribute choice have been reviewed by Krawczyk (2002); Venkatraman and Huettel (2012) and Volz and Gigerenzer (2012). These papers reviewed empirical evidence for the neural mechanisms of multi-attribute choice, however they did not focus on strategy use. Here, we focus on the use of compensatory and noncompensatory strategies, in situations when both kinds of strategies can be used, that is when knowledge beyond option recognition is available. Therefore, we did not include studies on the recognition heuristic by Volz et al. (2006) and Rosburg et al. (2011), because these studies analyzed situations when only recognition heuristic could be applied (or not) and did not look at the problem of selecting a strategy from a broader repertoire.

Several studies have attempted to identify the underlying neural sources of interindividual variability in strategy use. Venkatraman et al. (2009a), studying choices under risk using fMRI, showed that individual variability in the use of compensatory and simplifying decision strategies could be predicted by activity in ventral striatum, suggesting that high activity in the dopaminergic system might underlie the tendency to use the simplifying noncompensatory strategies. Moreover, Venkatraman et al. (2009b) provided evidence that strategic control of decisions is associated with the activity of anterior dorsomedial prefrontal cortex (DMPFC, see also Venkatraman and Huettel, 2012, for a review). Similarly, Gluth et al. (2013a), studying inference based multi-attribute choice with fMRI, showed the association between strategy selection and the activity of ventral striatum, the ventromedial prefrontal cortex and the anterior cingulate cortex. Khader et al. (2011) using a training paradigm and fMRI recording during a memory-based multi-attribute choice, showed that the amount of information that was required to use in a decision was reflected by the activity of left dorsolateral prefrontal cortex (DLPFC), and this activity modulated the activity in posterior areas associated with storage of decision cues. Replicating and extending these findings, Khader et al. (2015) showed that DLPFC activation is sensitive to both the number of decision cues that is retrieved in a controlled manner and the number of cues that are automatically activated by a decision option, suggesting that DLPFC activation reflects a general retrieval effort. Khader et al. (2011, 2015) also found that activation in the parietal cortex increased as a function the number of attributes to be retrieved. Together, these findings constitute important evidence for the involvement of the frontoparietal network in predecisional information processing.

The problem of using the right strategy for a particular decision task is framed as the strategy selection problem and several models have been proposed to account for this process (Table 2). The earliest are Beach and Mitchell (1978), Christensen-Szalanski (1978) and Payne et al.'s models (1988, 1993), which take a top-down approach to strategy selection. Regardless the differences, these models commonly assume that along with the repertoire of strategies, decision makers possess a priori knowledge of the cost and benefits of using a particular strategy, and integrating this knowledge leads to a (presumably) conscious, deliberate choice of a strategy. Strategy selection learning theory (SSL, Rieskamp and Otto, 2006), like the previous models, assumes that people possess a repertoire of strategies from which they can choose the appropriate strategy. It postulates that reinforcement learning of expectations of strategies' performance is the main driving force behind the tendency to use a particular strategy.

Another model is Lee and Cummins' Evidence Accumulation Model (EAM, Lee and Cummins, 2004), which assumes that both the rational decision strategy (RAT) and the fast and frugal TTB strategy are special cases of a sequential sampling decision process. Thus, it is a rather radical departure from the previous models, because the assumption of the repertoire of strategies is absent in this model. In turn, its main assumption is that the TTB strategy and the rational strategy can be unified within one process. Building on EAM, Bergert and Nosofsky (2007) also proposed that TTB strategy and the rational model can be unified within one framework. The crucial difference was that Begert and Nosofsky relaxed several assumptions of the earlier model in order to make it more psychologically plausible, e.g., they assumed that decision makers do not always use the optimal attribute weights and that the final choice is not made deterministically, but rather in a probabilistic manner. Glöckner and Betsch (2008) proposed a connectionist, parallel constraint satisfaction model of predecisional information processing. Similarly to the above models, their approach is also based on a unifying mechanism. Their model proposes that probabilistic decision tasks can be represented as simple neural networks (Glöckner and Betsch, 2008). It assumes that the activation spreading through such networks and eventually settling in a balanced state is the underlying mechanism of choice. The settled state of network activation represents the state of maximum consistency, which is compared to a threshold, and when the given threshold is exceeded, the choice is made.

In sum, the strategy selection problem was first approached from the perspective of top-down models, assuming a repertoire of strategies that are themselves selected by a meta-mechanism that takes into account situational factors and decision maker's resources. Departures from this line of reasoning were bottom-up models which attempted to explain the apparent variability in strategy use with unifying mechanisms. Together, these models offered some insights into the cognitive processes underlying decisions. However, they did not offer any mechanistic explanation of how the model parameters might be changed. In other words, these models were not well grounded in theories of elementary cognitive processes. This opens the way for more explicit accounts of how strategy selection is shaped. We believe that further unified models should be proposed, grounded in empirical and theoretical work on attention, and in the neurophysiological work on the neural substrates of attentional processes. Since the evidence for the neural underpinnings of strategy selection is growing, these models should be designed as neurocognitive rather than purely cognitive models. Moreover, since the evidence for the impact of emotions on strategy use is also growing (e.g., Scheibehenne and von Helversen, 2015; Wichary et al., 2015a) and neural theories of how emotions impact cognition are available (Aston-Jones and Cohen, 2005; Mather et al., 2015), such models should incorporate emotional factors and their underlying neural substrates to fill the existing gap in explanation of how emotions impact decision making.

In the following sections, we present a neurocognitive model of predecisional information processing and decision strategy selection. In order to go beyond empirical data and to understand the cognitive processes fully it is important to test mechanistic process models that try to explicitly state the underlying neurocognitive mechanisms (cf. Wichary et al., 2015b; Chuderski and Smolen, 2016). Within the domain of strategy selection in multi-attribute choice, there exists an empirical challenge described by Newell (2005) that it is impossible to distinguish, based on behavior, the accuracy of the unified and multiple strategy models described in the previous section. Therefore, we think it is time to build models of decision strategy selection that can provide predictions that go beyond behavior and that can be tested in neurophysiological studies. The need to develop such models is also underlined by the fact that several neurophysiological models of multi-attribute choice already exist (Louie et al., 2013; Chau et al., 2014; Hunt et al., 2014; Tsetsos et al., 2016). These models are relevant here, because they stress the importance of brain-wide computational processes as determinants of choice—particularly, the divisive normalization, hierarchical inhibition and gain control processes. Our model shares these computational principles and applies them to strategy selection.

The model, BUMSS (Bottom-Up Model of Strategy Selection), is based on our previous attempts to understand strategy selection in a unified, bottom-up manner (Smoleń and Wichary, 2008; Wichary and Smoleń, 2013). In these early version, as well as in the current version, the basic assumption is that in every choice situation people attempt to compute a weighted additive evaluation of decision alternatives. With this assumption, the question arises how to shape this general underlying process so that it resembles information processing with a simple noncompensatory heuristic at times. The idea behind our model is based on the observation by Martignon and Hoffrage (1999) that the TTB heuristic is equivalent to the WADD rule with a noncompensatory set of cue weights. A noncompensatory set of cue weights is a J-shaped distribution of cue weights, where one cue has a much higher weight than the other cues, such that the sum of their weights is not higher than the weight of the best cue. In contrast, a compensatory cue weight distribution is such that the cue weights are similar, and the sum of the less valid cues is higher than the weight of the best cue (Martignon and Hoffrage, 1999, 2002). In this perspective, the main question that our model tries to answer is how to obtain a noncompensatory set of cue weights using neurally plausible computations.

To illustrate the problem, consider a two-alternative, six-attribute choice task. The goal of the decision maker is to choose the more expensive diamond from the set of two. Each of the diamonds is described with six cues (attributes): size, overall proportions of the diamond, crown proportions, pavilion proportions, size of table and color (e.g., Mata et al., 2007). Each cue has a utility. In studies on multi-attribute choice, utility of a cue is often defined by the experimenter as cue validity—the conditional probability that a choice based on this cue is correct, given that the cue discriminates between the choice alternatives (Rieskamp and Hoffrage, 1999; Martignon and Hoffrage, 2002). There are many alternative measures that capture the utility of a cue, e.g., bayesian validity (Lee and Cummins, 2004), usefulness (Newell et al., 2003), success (Newell et al., 2004). Also, subjectively, cue utility can be determined by the decision maker through a host of reinforcement learning and memory processes that reflect how well the cue performed in the past and how relevant the cue is to the current goals set by the decision task (e.g., Rakow et al., 2005). The different definitions of cue utility and the processes that compute it are beyond the scope of the current paper. However, in many experiments on inference based multi-attribute choice, cue validity is explicitly given to the participants as a number, and participants use this information to guide their information acquisition and choice (e.g., Rieskamp and Hoffrage, 1999, 2008; Bröder, 2000, 2003; Newell et al., 2003; Newell and Shanks, 2003). In the current model, we stay with this conceptualization and assume that cue validities (Q) are available to the decision makers as the following numbers: 0.706 (size), 0.688 (overall proportions), 0.667 (crown proportions), 0.647 (pavilion proportions), 0.625 (size of table), and 0.6 (color). Each cue takes one of the two values for each diamond: low (0) or high (1) (see example in Table 3).

According to BUMSS, there are three processes that form the bottom-up mechanism of decision strategy selection and lead to the final choice: (1) cue weight computation, (2) gain modulation, and (3) weighted additive evaluation of alternatives (Figure 1). BUMSS postulates that in order to choose one alternative from several alternatives, the decision maker will first acquire and then weigh available decision cues, according to the softmax rule (Sutton and Barto, 1998; Doya, 2002), which is an example of divisive normalization (Louie et al., 2013): (1)aj=eβQj∑i = 1keβQi where a_j is the weight of decision cues j, Q_j is the cue validity of cue j and β is the gain control parameter. The crucial parameter in Equation (1), β (inverse temperature), reflects the changes in neural gain (e.g., Doya, 2002; Aston-Jones and Cohen, 2005). In computational neuroscience terms, gain is the amplification of neural activation, which is achieved by additional input into a signal processing pathway and is usually expressed as multiplication of the signal (Eldar et al., 2013; Louie et al., 2013). According to Equation 1, changes in gain lead to the changes in cue weighting (Figure 2). In our example, gain increase from 1 to 35 will transform the original cue validities (0.706, 0.688, 0.667, 0.647, 0.625, and 0.6) with a compensatory distribution into the resulting cue weights (0.501, 0.267, 0.128, 0.063, 0.029, 0.012) with a noncompensatory distribution (Table 3). Note that in our example such a noncompensatory distribution is only obtained with a very high value of β (35), whereas small values of β (1 or 5) lead to compensatory distributions of cue weights (Figure 2).

After acquiring and weighting the decision cues, the decision maker will then attempt to evaluate decision alternatives by integrating the weights and values of the decision cues, which is described by the weighted additive rule: (2)Vj =∑i = 1kai ci,j where V_j is the value of a choice alternative j, a_i is the cue weight and c_{i, j} is the value of ith cue for jth alternative. In our example, the option values are V₁ = 0.335, V₂ = 0.474 for β = 5 and V₁ = 0.513, V₂ = 0.342 for β = 35, thus showing that different values of β parameter can lead to qualitatively different option evaluations. Finally, the decision maker will choose an alternative j from several alternatives in a probabilistic manner, with a probability d_j that is given again by the softmax rule (Sutton and Barto, 1998; Doya, 2002): (3)dj = eVjβ∑i = 0oeViβ where V_j is the value of alternative j, o is the number of options and β is the same inverse temperature parameter as in (1) that reflects the gain and in consequence, controls the stochasticity of the choice. For increasing β, the probability that the alternative with the highest value will be chosen, increases. For a choice between two alternatives V₁ and V₂, this will be shown as changes in the steepness of the well-known sigmoid function (Figure 2). In our example, with the low values of β (in both Equations 1 and 3), the probabilities for alternative 1 and 2 will be equal to 0.333 and 0.667, respectively. With high values of β, the probabilities will be equal to 0.997 and 0.003, respectively. Note the influence of β on this stage. Although the absolute difference between the values of the alternatives computed by Equation (2) was similar in both cases (0.14 and 0.17) the probabilities of choice differ substantially for different values of β.

Apart from specifying the computational processes, the aim of the current model is to propose a set of neural substrates underlying choices consistent with WADD and TTB. We postulate that the computational processes described above take place in a brain network consisting of (1) the dorsolateral prefrontal cortex (DLPFC) and parietal cortex (PC)—which form the fronto-parietal network (FPN) and are responsible for cue weight computation and weighted additive evaluation, (2) the orbitofrontal cortex (OFC), which contributes to cue weight computation, (3) the anterior cingulate cortex (ACC), and (4) the brainstem nucleus locus coeruleus (LC), which contribute to cue weight computation and choice by modulating gain (Figure 1).

Fronto-parietal network, particularly its components the dorsolateral prefrontal cortex (DLPFC) and the parietal cortex (PC), performs the operations of maintaining and integrating information necessary for current and future behavior and thus is the neural substrate for working memory processes (Linden et al., 2003; Erikson et al., 2015). Orbitofrontal cortex (OFC) computes positive and negative utility of incoming stimuli (Tremblay and Schultz, 2000; Aston-Jones and Cohen, 2005; Padoa-Schioppa and Assad, 2006; Wallis, 2007; Volz and von Cramon, 2009). Anterior cingulate cortex (ACC) serves a broad range of functions (Paus, 2001) such as conflict monitoring, cognitive control (Botvinick, 2007), and regulation of autonomic activity (Critchley, 2005). Growing evidence suggests that it computes negative utility or cost of the undertaken actions (Aston-Jones and Cohen, 2005). Although evidence suggests that OFC and ACC may both code positive and negative utility, within the current model, we stay with the Aston-Jones and Cohen's (2005) simplifying assumption that the OFC codes the positive utility of incoming stimuli and the ACC encodes the negative utility of engaging in a given behavior.

ACC receives multiple inputs from various brain structures: prefrontal cortex, parietal cortex, amygdala, insula and somatosensory cortex, among others (Paus, 2001). Inputs from the PFC relay the information on the information processing demands associated with performing a current task (Paus, 2001). Inputs from the parietal cortex, mainly the temporo-parietal junction (TPJ), relay the information about stimuli salience, therefore the ACC (together with the TPJ, insula and locus coeruleus) is involved in computing cue salience (Menon and Uddin, 2010; Litt et al., 2011; Kahnt and Tobler, 2013; Kahnt et al., 2014). In addition, inputs from the amygdala, insula and somatosensory cortices relay information about threatening or noxious stimuli to ACC (Bornhövd et al., 2002; Kuo et al., 2009; Etkin et al., 2011). It is important to note that ACC activity is modulated by a wide range of experimentally manipulated variables, including pain, threatening stimuli, conflicting cues, physical and cognitive effort (Shackman et al., 2011) and thus can serve as an integrated representation of the noncognitive context of decision processes.

Brainstem nucleus locus coeruleus is involved in regulating arousal and attention, the sleep-wake cycle and the stress response. It is a main source of brain norepinephrine (NE). LC neurons respond in a tonic and phasic manner to aversive as well as appetitive stimuli and release NE that mediates the mobilization of energy for avoiding aversive stimulation or obtaining a reward (Rajkowski et al., 2004; Aston-Jones and Cohen, 2005; Sara and Bouret, 2012). Importantly, NE release regulates the gain of information processing in the cortex. LC performs this by widespread connections with most of the brain. It has reciprocal connections with the dopaminergic system that implicate LC in mobilization for reward processing (Varazzani et al., 2015). Both OFC and ACC have strong reciprocal connections with LC (Aston-Jones and Cohen, 2005) and OFC may relay its information about stimulus utility via ventral striatum (Ullsperger et al., 2014). LC is connected to FPN, where it projects its diffuse connections which are the basis for cortex-wide gain modulation (Eldar et al., 2013). LC is also reciprocally connected to amygdala, which regulates LC function by corticotrophin releasing hormone (CRH, Owens and Nemeroff, 1991). Similarly, LC is reciprocally connected to hypothalamus (and thus to the whole hypothalamic-pituitary-adrenal axis), which activates it by CRH (Owens and Nemeroff, 1991). Through a common afferent, the nucleus paragigantocellurlaris, LC is also connected to the sympathetic-adrenal-medullary (SAM) system and thus it is co-activated with the peripheral effectors of the autonomic nervous system responsible for the galvanic skin response or pupil dilation (Nieuwenhuis et al., 2011).

The aim of this paper is two-fold: first, to review the existing evidence on decision strategy use, the cognitive models and the neural underpinnings of decision strategy selection and second, to propose an original neurocognitive model of this process, with a particular focus on the role of noncognitive factors in shaping strategy selection. To this end, we reviewed the relevant literature on strategy selection and identified gaps in our knowledge of factors influencing this process. Particularly, only a limited number of studies took into account the impact of emotional factors on predecisional information processing and decision strategy use. These gaps suggest that further studies are needed in this domain. The review of the existing models of decision strategy selection also pointed out that most of the existing models are top-down, that is they assume a master, metacognitive process that selects among the decision strategies available to the decision maker. This feature of the top-down models leads to one well known problem: the need for a metastrategy to select among the individual strategies which eventually leads to the regress to infinity (e.g., Newell, 2005). Another serious problem of the top-down frameworks is strategy sprawl, that is constant adding of new heuristics to the repertoire (Scheibehenne et al., 2015).

Such limitations led to the formulation of unifying models which attempted to reduce the multiplicity of heuristics to more general information processing mechanisms. Our model presented in this paper is another proposal how to unify the use of the rational, compensatory decision strategies with the noncompensatory, one-reason decision heuristics. It is closest to Lee and Cummins (2004) and Bergert and Nosofsky (2007) models, however it goes beyond these models. This model attempts to reduce the apparent variability of decision strategies by offering a unifying mechanistic account that is rooted in neurophysiology. Thus, it offers a mechanistic neural processing account of decision strategy selection. Moreover, it allows for incorporating noncognitive factors that may influence strategy selection, such as emotional arousal, stress and effort. It does so by explicitly linking higher level cognitive processes that take place in cortical areas with brainstem activity, and shows how this activity can shape decision strategies and eventually, choices. Because the territory of how emotions impact decision making is rather uncharted, our model also offers some of the first mechanistic insights how these processes might be operating.

SW and TS developed the theoretical model, and wrote and edited the manuscript.

This research was supported by the postdoctoral fellowship from the Copernicus Centre for Interdisciplinary Studies in Krakow, an intramural Grant No. WP/2015/A/25 from the University of Social Sciences and Humanities in Warsaw and a grant from the Ministry of Science and Higher Education of Poland (NN106219338) to SW.