Special Reviews

Anterior Prefrontal Function and the Limits of Human Decision-Making

See allHide authors and affiliations

Science  26 Oct 2007:
Vol. 318, Issue 5850, pp. 594-598
DOI: 10.1126/science.1142995


The frontopolar cortex (FPC), the most anterior part of the frontal lobes, forms the apex of the executive system underlying decision-making. Here, we review empirical evidence showing that the FPC function enables contingent interposition of two concurrent behavioral plans or mental tasks according to respective reward expectations, overcoming the serial constraint that bears upon the control of task execution in the prefrontal cortex. This function is mechanistically explained by interactions between FPC and neighboring prefrontal regions. However, its capacity appears highly limited, which suggests that the FPC is efficient for protecting the execution of long-term mental plans from immediate environmental demands and for generating new, possibly more rewarding, behavioral or cognitive sequences, rather than for complex decision-making and reasoning.

Decision-making is the realm of the frontal lobes, ranging from the simplest choice, for example, choosing an apple or a pear from a fruit basket, to the most complex ones like deciding the next move in a game of chess. Decision-making is subserved by a hierarchy of lateral frontal cortex regions controlling action selection in relation to internal drives and subjective preferences, the apex of which corresponds to the lateral convexity of the cyto-architectonic area defined as Brodmann's area 10, the so-called frontopolar cortex (FPC) (Fig. 1A) (1). The FPC is tremendously well developed in humans compared with other primates (2), and brain imaging has provided evidence that it contributes to uniquely human traits like reasoning or problem-solving. However, the FPC is not a homuncular command center responsible for orchestrating cognition in lower brain regions. Patients with FPC lesions show no significant impairments on formal neuropsychological tests of perception, language, and intelligence but appear markedly impaired in decision-making in open-ended and ill-structured situations, which often occur in everyday life (3). These observations suggest that the FPC has evolved as a functional “add-on,” perhaps to overcome the limitations of more posterior prefrontal processes by enabling the emergence of more flexible cognitive control in the service of decision-making.

Fig. 1.

The human frontopolar cortex. (A) The FPC corresponds to the lateral Brodman's area 10 (red region), the most rostral portion of the human prefrontal cortex [from (35)]. (B) FPC activation observed with functional magnetic resonance imaging (fMRI) on a horizontal brain slice (dashed line indicates approximate localization) in multitasking behavior, in which subjects postponed the execution of a task to perform another first [from (27)]. (C) fMRI also revealed that FPC activity correlated with the amount of uncertainty (i.e., entropy) remaining between multiple putative behavioral options that subjects were simultaneously tracking while exploring a virtual maze [from (8)].

The characterization of FPC function is a major issue in current cognitive neurosciences, not least because it has the potential to reveal an upper bound to human executive function. Here, we review recent brain imaging studies that provide important new insights into FPC function. The FPC enables contingent interposition of multiple mental tasks or behavioral plans, the function that allows humans to overcome the serial constraint upon control processes evident in more posterior prefrontal regions (4). We describe a simple neurocomputational model delineating the basic computational properties of this FPC function and explaining its mechanistic implementation in neuronal interactions between FPC and neighboring prefrontal regions.

FPC Contributions to Human Cognition

Learning and exploration. The FPC is robustly engaged when subjects are instructed to learn new behavioral routines (57). This engagement is observed in both reinforcement (5, 7) and supervised (6) learning paradigms, indicating that the FPC contributes to explicit learning regardless of the type of learning or feedback. However, in contrast to more posterior prefrontal regions, the FPC gradually disengages over the course of learning (6). This disengagement occurs as subjects switch back and forth between alternative behavioral options (exploration), drawing upon feedback to progressively eliminate irrelevant choices. Indeed, to a greater extent than other prefrontal regions, FPC activity specifically correlates with the amount of uncertainty (i.e., entropy) remaining between multiple putative options that subjects are simultaneously tracking (8) (Fig. 1C). Moreover, FPC is active whenever subjects depart from an a priori optimal option to check alternative ones (9). Thus, the FPC contribution to learning and exploration appears to be associated with maintaining and switching back and forth between multiple behavioral alternatives in search of optimal behavior.

Memory retrieval. The FPC is also engaged in episodic memory tasks, in which subjects are instructed to judge whether multiple serially presented stimuli have been previously encountered in specified past episodes (10). The FPC is generally involved in such episodic retrieval paradigms regardless of the nature of past episodes and stimuli, retrieval success, or stimulus novelty (11). The FPC is tonically active during retrieval (12), in preparation for subsequent retrieval (13) and also in response to probe stimuli that occur less frequently (14). The involvement of the FPC thus cannot be accounted for by memory retrieval per se but is likely to be accounted for by the task structure requiring scheduled retrieval of multiple elements of a past episode. This view has recently received direct experimental support: FPC activity in episodic retrieval tasks is virtually identical to that underlying categorization tasks that require a combination of multiple cognitive subtasks (15, 16). Thus, FPC activations observed in episodic retrieval paradigms are likely to result from the coordination of subtasks involved in those paradigms, including the recurrent retrieval of multiple elements of past episodes and the demands of the judgment task pertaining to each stimulus.

Relational reasoning. The FPC is involved in relational reasoning, in which subjects are required to integrate the outcomes of multiple inferences for selecting appropriate responses. This has been observed for various types of inferences, including perceptual (17, 18) and semantic (19) inferences. More generally, the FPC is engaged whenever subjects must integrate the outcomes of multiple internal subtasks, for example, when deciding whether multiple items share the same property (15, 16) or, conversely, whether a single item jointly exhibits multiple properties (20). Is the FPC active only at the final integration stage, when multiple outcomes are combined to reach a decision (10), or does the FPC underlie the passing of information between successive computational stages in the service of eventual integration? Recent experimental results rule out the former hypothesis (21), because FPC activity was observed even in the absence of a final integration stage, and the FPC exhibited sustained activations in the intermediate phase preceding final integration. Instead, in relational reasoning, the FPC allows the buffering of information from previously executed subtasks while subsequent stages are being carried out.

Multitasking behaviors. Finally, the FPC is robustly engaged in multitasking behaviors, in which subjects postpone the execution of one task to perform another first (22). Multitasking recruits the most anterior regions in the FPC (23) (Fig. 1B) and is selectively and severely altered in patients with FPC lesions (24). Moreover, unlike other prefrontal regions, the FPC involvement in multitasking is not reducible to component processes such as task-switching or task-delaying (22). The FPC specifically exhibits sustained activations associated with postponed tasks that resist distraction and ongoing performance regardless of the nature of tasks (25, 26). Task contingency underlying multitasking behaviors is also a key factor in activating the FPC, given that it is not engaged during the execution of structured superordinate plans composed of multiple embedded tasks (27). Overall, these findings indicate that the FPC specifically subserves the ability to contingently switch back and forth between independent tasks by maintaining distractor-resistant representations of postponed tasks during the performance of another.

The Core Function of the FPC

The lateral prefrontal regions are involved in selecting and maintaining action selection rules (i.e., task sets) according to the immediate context and/or the ongoing temporal episode in which the person is acting (28). This lateral prefrontal system obeys a serial principle, which allows only a single task set to govern action selection at any one time (4). Such a serial constraint minimizes conflict situations that may arise between multiple task sets and thus optimizes action selection. The FPC overcomes this serial constraint by enabling the joint consideration of multiple task sets in a wide variety of behavioral paradigms. Collectively, the data depicts an anterior prefrontal system in which lateral prefrontal regions select and maintain the task set governing ongoing action, whereas the FPC enables previously selected task sets to be maintained in a pending state for subsequent automatic retrieval and execution upon completion of the ongoing one. This process, which we call “cognitive branching,” forms a domain-general core function at the basis of the behaviors and mental activities requiring simultaneous engagement in multiple tasks that are not serially organized into a single, pre-established superordinate plan (27).

Accordingly, the branching process enables a task or a behavioral episode to be temporarily suspended while another is being performed, or conversely allows reversion to a pending task or episode following completion of the ongoing one, even in the absence of any external cues. Such an ability is especially impaired in patients with frontopolar lesions (24). Moreover, the branching process is critical for relational reasoning, because the maintenance of pending task sets associated with previously computed inferences during subsequent inferential tasks (and reverting back to such pending representations) is required for integration. For example, in the calculation (2 + 4) × (5 + 7), the first addition is computed (= 6), then the operand/task-set 6 is maintained in a pending state while the second addition is computed (= 12). Then, reverting back to the pending operand will cue the multiplicative operation 6 × 12, leading to the final result (= 72) (21). The branching process is also critical for exploration and tree-search because it allows switching in and out of “branches” or behavioral options while maintaining others that are pending (9) (8). Although episodic memory retrieval presumably does not require branching processes, it is likely that cognitive branching is engaged in a series of multiple, concurrent retrieval trials. The branching process may be involved in performing the judgment task pertaining to each trial while maintaining specified past episodes in a pending state so as to internally retrieve them subsequently for preparing the next memory judgment trial. Also, episodic or “source” memory may require subjects to hold data in a pending state and switch between the putative episodes with which the retrieval probe could have been associated. Finally, compared with more posterior prefrontal processes, the key feature of branching processes is to enable resumption of pending task sets in the absence of any external cues or associations between the ongoing and pending task sets, so that the execution of pending task sets is only triggered by termination of the ongoing task. Consequently, when contingency, uncertainty, or entropy over task sets decreases—for example, when serial associative links develop among several task sets because of training or simply when the number of possible task sets shrinks—the need for branching processes decreases and the FPC is expected to disengage. In agreement with this hypothesis, the FPC disengages when contingency, uncertainty, or entropy over multiple task sets decrease during learning (6), exploration (8), or multitasking (27) (Fig. 1C). Thus, the branching hypothesis accounts for the FPC involvement in a wide variety of behavioral paradigms, which suggests that cognitive branching forms core function of the FPC. This function substantiates the more phenomenological view that the FPC plays a pivotal role in switching between externally versus internally oriented thoughts (3) (supporting online text).

A Neurocomputational Model of FPC Function

It remains to be described how the anterior prefrontal system “decides” to place an ongoing task into a pending state and to revert back to it later. This is a key theoretical issue, given that the FPC is not under the control of higher brain centers. The FPC is specifically engaged when subjects suspend the execution of an ongoing task set associated with a priori the largest expected future rewards in order to explore a possibly more rewarding task set (9). The result suggests that with no supervisory optimization, cognitive branching occurs between two concurrent behavioral options, when reward expectations associated with each option (or expected penalties if not executed) are large enough so that it would be too costly or risky to simply abandon one. In that event, comparing the rewards expected from executing each option immediately will determine which option is placed in a pending state (the less rewarding one) and which one is selected for guiding immediate behavior (the more rewarding one). We proposed a minimal neurocomputational model showing how the FPC (denoted Fpc) may mechanistically implement reward-based cognitive branching with no supervisory optimization through interactions with neighboring prefrontal regions, namely, the anterior lateral and medial/orbital prefrontal regions (denoted Lpc and Ofc, respectively) (29) (Fig. 2 and supporting online text). In accordance with empirical results (28, 30, 31), the model assumes that Lpc neurons represent the active task set guiding current behavior, that is, active Lpc units through top-down projections select and maintain task-set specifications represented in other brain regions and required for execution (typically caudal prefrontal regions), whereas Ofc neurons code for the future rewards expected from executing task sets. Lpc and Ofc neurons are reciprocally connected and receive input signals from other brain regions (typically posterior associative and paralimbic cortices) cueing activity toward specific task sets and updating expected future rewards. By contrast, Fpc neurons are reciprocally connected to Lpc and Ofc neurons only, with an inhibitory influence of Lpc onto Fpc neurons. Overall, according to reward expectations, Fpc units form a possible back-up buffer for storing a previously selected task set in Lpc, while Lpc units are representing another. Subsequently, the Fpc buffer enables to possibly reinstantiate in Lpc the pending task set and consequently to reinstantiate its specifications in lower brain regions for execution, even in the absence of any external inputs.

Fig. 2.

A neurocomputational model of the frontopolar function. Octagons represent neuronal populations in the lateral prefrontal (Lpc), medial/orbital frontal (Ofc), and frontopolar (Fpc) cortex. Green and red arrows indicate excitatory and inhibitory interactions, respectively. Strong lateral inhibition within Lpc and Fpc enforces task-set selection; strong self-excitation (not shown) enables task-set maintenance in the absence of any external inputs. By contrast, weak lateral inhibition and self-excitation within Ofc enables Ofc neurons to maintain reward expectations related to multiple task sets, but only those corresponding to the task sets maintained in Lpc and Fpc. All interactions between Ofc, Lpc, and Fpc neurons are excitatory except the influence exerted by Lpc on Fpc neurons, which is assumed to be inhibitory. The inhibitory influence of Lpc onto Fpc neurons forces Lpc and Fpc to encode distinct task sets. Ofc also includes an input layer storing and updating expected reward values associated with task sets with respect to input signals (e.g., external or internal contextual cues, feedback, and task end-states). Neuron activity is governed by standard activation dynamics equations (supporting online text).

Computer simulations show that this neuronal system forces Lpc and Fpc neurons to potentially select and maintain only the two most rewarding task sets. The other task sets are discarded. Cognitive branching occurs between the two most rewarding task sets, provided that the second largest expected reward is larger than a given threshold Rb (Fig. 3 and fig. S1). The most and second-most rewarding task sets are then selected and encoded in Lpc and Fpc neurons as the active and pending task set, respectively (Fig. 3, C and D). Consistent with experimental data (25), Lpc and Fpc neurons show sustained activation during the pending period. This situation perpetuates (or possibly reverses according to updated relative reward values) until one expected reward value drops below threshold Rb (e.g., because of task completion). If this value corresponds to the active task set, Lpc neurons start encoding the pending task set and Fpc activity returns to background noise level: the active task set is discarded and the pending task set becomes the active one for guiding subsequent behavior, thereby achieving a branching between the two task sets (Fig. 3C). Conversely, if the value corresponds to the pending task set, this one is discarded from the Fpc and branching is aborted (Fig. 3D). Also, no branching occurs if the second-largest expected reward does not reach the reward threshold Rb (Fig. 3, A and B). In that event, no sustained activity occurs in Fpc neurons during the execution of the most rewarding task set. The second-most rewarding task set is discarded, and the system does not revert back to it in the absence of any subsequent external cueing. However, consistent with experimental data (32), Fpc neurons may exhibit short phasic responses to external cues associated with multiple task sets (Fig. 3D).

Fig. 3.

Phase diagram and activation dynamics in the frontopolar model. Task sets are cued by an external cue, C, at time 0. (Left) Phase diagram showing different model behaviors according to the reward value (R1 and R2) expected from the two most rewarding task sets, T1 and T2, once cue C occurs. R1 and R2 remain unchanged until the occurrence of feedback signals X1 and X2, respectively. The phase diagram exhibits two reward thresholds, Ra and Rb, delimiting five regions with distinctive behaviors. The phase diagram is symmetrical, so that we describe only the regions where R1 > R2. (Right) (A, B, C, and D) show typical time courses of Fpc, Ofc, and Lpc neuron activity for each region corresponding to values R1 and R2 shown on the phase diagram (points A, B, C, and D in the left panel). Horizontal and vertical axes represent time and neuron activity, respectively. Blue and red lines show neuron activity coding for T1 and T2, respectively. (Purple region) R1 and R2 are not large enough (<Ra) for the network to select and encode any task sets. No tasks are executed. (Blue region) R2 is not large enough (<Rb), and no branching occurs. Only the most rewarding task set, T1, is encoded in Lpc for guiding current behavior (active status). All other task sets are discarded. In (A) and (B), expected reward R1 drops below Ra once feedback X1 occurs, terminating T1 execution. T2 remains unselected after T1 termination, although R2 becomes larger than R1 (a new external cue would be required for selecting T2). Note in (B) the phasic Fpc response to cue C, because R1 and R2 are close to threshold Rb. (Cyan region) R2 is large enough (>Rb) so that T2 is encoded in Fpc and placed in a pending state, whereas T1 is encoded in Lpc for guiding current behavior. (C) Expected reward R1 drops below Rb once feedback X1 occurs, whereas R2 is still above Rb. T1 is terminated and automatically replaced by T2 in Lpc for guiding subsequent behavior. Cognitive branching has occurred. (D) As in (C), except that R2 drops below Rb before R1 (X2 occurs before X1). As a result, the pending task set, T2, is discarded from the Fpc. T1 remains the active task set guiding behavior. Cognitive branching is aborted.

In sum, the second most rewarding task is placed in a pending state in the FPC, while lateral prefrontal regions are controlling the execution of the most rewarding task, provided that the future rewards expected from the two task sets are large enough (>Rb). Thus, cognitive branching occurs and FPC is recruited with no supervisory optimization and control, because it would have been too costly to discard the second most rewarding task as suggested by empirical results (9). The model indicates that the reward threshold Rb for branching and recruiting FPC is larger than the minimal expected reward (Ra) required for triggering task execution and recruiting lateral prefrontal regions (Fig. 3 and fig. S1). The model also shows that reward expectations associated with the active and pending task sets are continuously updated with respect to feedback signals related to current behavior. In particular, if reward expectations associated with the pending task set drop below the reward threshold Rb for branching, cognitive branching is aborted and the pending task set is discarded. This predicts that the anterior prefrontal system enables online evaluation of the pending task set according to the outcomes of a distinct, ongoing task, an ability referred to as the valuation of fictive action that may considerably speed up learning and exploration (33).

This model does not aim to describe in detail neural coding of task sets in the anterior prefrontal regions. Nevertheless, it explains how the processing of cognitive branching may mechanistically emerge from neuronal interactions between FPC and neighboring prefrontal regions. The model provides a predictive theoretical framework within which we can empirically address several unresolved major issues about the FPC function and its relationships with other prefrontal functions, including the processing of internal drives and subjective preferences in the medial and orbital sectors of the prefrontal cortex (30, 31).

In particular, a key prediction is that the capacity of the FPC cannot exceed the processing of a single pending task at any one time: According to the model, the anterior prefrontal system can process only the two most rewarding task sets, even if they are associated with identical expected reward values. Interferences supervene in Fpc, when three or more distinct task sets are associated with the two largest expected reward values, in which case cognitive branching is dramatically impaired (supporting online text, fig. S2). Thus, the model especially predicts that the FPC is unable to recursively perform cognitive branching—resuming a primary and secondary pending task after completion of a third task—because interferences supervene between the two pending tasks. This prediction has been recently tested in our laboratory in a behavioral protocol that recruits the FPC and requires subjects to engage in multitasking behaviors. Our data confirm the prediction. We observed that compared with various control conditions, postponement of a second task to execute a third task while a first was already pending led to marked selective impairment in subsequently resuming the first and second pending tasks. Thus, the model and these preliminary results support the idea that the prefrontal executive system lacks the computational power to perform recursive cognitive branching, and consequently to control recursive tree-searches in the exploration of deep branching sets of future possible situations underlying reasoning, problem-solving, or complex decision-making.

This predicted, rather limited capacity of the FPC may appear surprising, given that humans can indeed carry out such complex decision-making. However, a complex decision situation may become tractable as expertise with a particular situation is acquired, reducing uncertainty and contingency during exploration. One possible hypothesis is that extensive training leads to the formation of mental spatial maps of branching sets, so that with expertise, recursive tree-searches reduce to spatial navigation, which relies on specialized brain structures like the parietal cortex and hippocampus. Alternatively, the property of recursion is a key feature of human language (34), and it is conceivable that expertise may lead to the formation of specific language-like coding systems mapping particular tree-structures and allowing to track complex and recursive exploration. Thus, the FPC capacity limit described above implies that complex decision-making in well-established situations does not critically rely on FPC. Consistently, frontopolar patients show few decision-making deficits in well-established situations such as standardized tests of intelligence, although they are impaired in open-ended and ill-structured situations (3).

Concluding Remarks

On the basis of recent empirical findings and neurocomputational mechanisms, we suggest that the processing of cognitive branching based on reward expectations with no supervisory optimization forms the core function of the anterior prefrontal cortex. This function partially overcomes the serial constraint that bears upon the control of mental tasks in the lateral prefrontal cortex, providing an additional degree of flexibility to the prefrontal executive system, which may be critical for the emergence of human higher cognitive abilities such as reasoning and problem-solving. However, we suggest that the FPC function is restricted to the processing of simple cognitive branching, whereby only a single task can be maintained in a pending state at any one time. This hypothesis places severe serial and recursive constraints on human reasoning, problem-solving, and complex decision-making. Consistent with this view, it appears unlikely that the human brain has evolved to solve complex problems such as deciding the next move in a game of chess. Selective pressure to survive in a physically challenging environment may place other demands before the need for such higher cognitive faculties. Nevertheless, a capacity-limited FPC function may have endowed humans with two key adaptive advantages: on the one hand, an ability to pursue long-term behavioral plans and at the same time respond to demands of the physical or social environments; on the other hand, to explore any potential gain from the interposition of new task sets within ongoing behavioral routines or from the contingent recombination of previously established behavioral plans, as in genetic recombination mechanisms. Thus, the frontopolar cortex may have played an even more critical role in the gradual formation of complex behavioral and cognitive routines such as tool use in individuals and societies, that is, in human creativity rather than complex decision-making and reasoning.

Supporting Online Material


SOM Text

Figs. S1 and S2


References and Notes

Stay Connected to Science

Navigate This Article