Emotion regulation from an action-control perspective

Despite increasing interest in emotional processes in cognitive science, theories on emotion regulation have remained rather isolated, predominantly focused on cognitive regulation strategies such as reappraisal. However, recent neurocognitive evidence suggests that early emotion regulation may involve sensorimotor control in addition to other emotion-regulation processes. We propose an action-oriented view of emotion regulation, in which feedforward predictions develop from action-selection mechanisms. Those can account for acute emotional-action control as well as more abstract instances of emotion regulation such as cognitive reappraisal. We argue the latter occurs in absence of overt motor output, yet in the presence of full-blown autonomic, visceral, and subjective changes. This provides an integrated framework with testable neuro-computational predictions and concrete starting points for intervention to improve emotion control in affective disorders.

Navigating personal and societal challenges requires proficiency in the regulation of emotional responses, both in terms of physiology and subjective experiences (feelings) as well as behavioral output (Hare, 2017;Koch et al., 2018).Consider the case of a socially anxious teenager encountering the person of their romantic dreams.The potentially appetitive encounter will elicit rapid physiological responses such as hormonal-and heart rate changes and the desire to increase proximity ('felt' action tendency to approach (Frijda, 1986)).However, predictions of social rejection might also trigger another Pavlovian (see glossary) action tendency, namely to flee or avoid, e.g. an urge to look or walk away from the romantic target to prevent disappointment.The desire to approach or avoid in order to achieve a certain outcome, and the progress made towards that goal, may be core to the subjective experienceand reflect the valenceof an emotion (Bach and Dayan, 2017;Frijda, 1986;James, 1890).Acting on the automatic avoidance tendency prevents the teenager from fulfilling the appetitive potential of that situation, e.g.finding out whether the romantic dreams are reciprocated.Crucial for the understanding of emotion regulation as action-selection, is the notion that neither inhibition of avoidance itself, nor ruminating about the missed opportunity, will be sufficient to achieve appetitive results.Instead, pursuing the appetitive potential of that situation (e.g.getting a date) involves selecting an alternative action, based on forward modelling of alternative action-possibilities and their hypothesized outcomes.
In this manuscript, we argue that action selection ("what if I act differently?")might underlie the full continuum of emotion regulation implementations, ranging from control over acute emotional or Pavlovian action tendencies to more abstract -model based -instances of emotion regulation such as cognitive reappraisal episodes (Fig. 1).Below, we articulate this action-oriented perspective on emotion regulation, considering three main arguments.First, a comparative argument, detailing emotion as an extension of negative feedback control processes that originally developed because they maintained homeo-and allostasis, and of which the input states are controlled by taking action (Cisek, 2019;Seth, 2015).Second, we consider a theoretical argument, namely the notion that forward modelling of actions and their outcomes is grounded in the pre-motor system.Importantly, this forward modelling of actions can be detached from overt motor output (Buzsáki et al., 2014;Pezzulo, 2011;Pezzulo and Cisek, 2016), but nevertheless modulate physiological and visceral reactions, as well as subjective components of emotion through corollary discharge (Crapse and Sommer, 2008;Pine et al., 2021).Third, we consider empirical evidence indicating that emotion-eliciting situations as well as emotion regulation consistently increase activity in sensorimotor and pre-motor systems.After detailing these three lines of evidence, we will end with specifying a model in which emotional-action control and emotion regulation are grounded in the extended pre-motor system, including much of prefrontal cortex (Fine and Hayden, 2022;Pine et al., 2021).This notion has important implications for understanding emotion regulation in clinical practice as will be discussed in the concluding section.

Emotion regulation needs action
Theoretical work has often framed emotion as a 'driver of action readiness' resulting from relatively rapid action-selection algorithms (Bach and Dayan, 2017;Frijda, 1986;Ledoux, 2007;Ridderinkhof, 2017).Crucial in this view is appraisal, the process of assessing the relevance of a situation in terms of the organism's goals and of the affordances the current situation provides (Moors et al., 2013).This view can explain many instances of emotion, including control over acute action tendencies elicited by homeostatically challenging situations (Mesquita and Frijda, 2011;Vilarem et al., 2020) and allostatic controllers that implement action to proactively influence internal signals (Sennesh et al., 2021;Seth, 2015).Also, it fits with recent evidence of motor-related neurophysiological activity during emotion-processing and emotional action control (Moors et al., 2019).However, this view has not been at the center of the most influential theories of emotion regulation (Gross, 2015a;Tamir, 2021).Those have traditionally, at least in their operationalizations, focused on cognitively or verbally-mediated emotion regulation strategies (McRae and Gross, 2020).These strategies include distraction and reappraisal of emotional content.They are primarily aimed at changing subjective feelings or physiological responses, less so on behavioral change (Etkin et al., 2015;Gross, 2015a;Kalisch et al., 2019;Ochsner and Gross, 2005;Tamir, 2021), but see (Hartley and Phelps, 2010).However, the fact that verbally accessible feelings may ultimately change as a consequence of emotion control does not imply that only verbal strategies underlie this achievement.Rather, comparative, theoretical, as well as recent empirical evidence point at mechanisms that allow individuals to select alternative emotional actions, mechanisms that developed ancillary to action selection through the extension of pre-motor systems (Fine and Hayden, 2022;Pine et al., 2021).This does not imply that subjective feelings should be banned from the field of emotion control.On the contrary, we aim to provide a phylogenetically and neurobiologically plausible way to explain the mechanisms underlying the full spectrum of emotion control instances.
In this work, we will extend action-oriented emotion theory (Bach and Dayan, 2017;Moors et al., 2019;Ridderinkhof, 2017) to a theory of emotion regulation that accounts for the arbitration of discrepancies between model-free or Pavlovian and model-based strategies.Our framework is based on action-outcome prediction, where actions can be employed to resolve discrepancy between actual and desired states (Barrett, 2017;Pezzulo and Cisek, 2016;Seth, 2015), thereby providing neurocomputational specificity to contemporary emotion regulation theories advocating a cybernetic perspective on emotion regulation (Gross, 2015a(Gross, , 2015b)).The prediction process involved is an instance of forward modeling, employing innately programmed or learned action-outcome relationships (Etkin et al., 2015) that can select between Pavlovian action-tendencies such as freeze-fight-flight reactions, as well as between instrumental actions, such as to approach or to avoid in order to achieve a certain goal (Fig. 1).We argue that the same prediction process can also explain forms of emotion regulation such as re-appraisal of emotional situations.Namely, this prediction process can be implemented through hierarchically nested control loops developed for action selection in pre-motor systems (Fine and Hayden, 2022;Pezzulo and Cisek, 2016) (Fig. 2; Key figure).This form of supervised forward modelling can change action selection by modifying goals and motivations, elicit changes in visceral and physiological responses and, subsequently, subjective components such as feelings, through prediction and corollary discharge of the action plans (Barsalou, 2008;Buzsáki et al., 2014;Winkielman et al., 2018).Based on these concepts, we argue that emotion regulation can be seen as a continuum of action-control, ranging from controlling automatic or Pavlovian (re)actions to more abstract instances of emotion regulation.
This contribution adds to current cybernetic theories of emotion control (Gross, 2015a;Tamir, 2021) in at least three manners: First, in our framework, emotion regulation does not aim to change the subjective state directly.Rather, emotion regulation is implemented as a consequence of action selection processes that serve to restore homeo-or allostasis, and the subjective feeling changes as a consequence of overt or modelled action.Second, current cybernetic theories of emotion control (Gross, 2015a(Gross, , 2015b;;Tamir, 2021) remain silent about implementation-level constraints.In contrast, the current contribution is informed by the neural implementation of the control mechanisms (e.g.Fig. 2).Third, we deem such detailed action-selection view of emotion regulation relevant for guiding interpretation of neural findings, and helpful to constrain the operationalization of emotion-regulation.

Fig. 1.
Emotion regulation from an action control perspective.When you find yourself in a challenging situation, for example encountering your secret beloved, you may choose to move away in order to avoid negative predicted outcomes that are associated with approaching to ask for a date (a process relying on unsupervised forward modelling).This first sweep of forward modelling is a phase that some researchers call 'appraisal' -leading to a Pavlovian action tendency.When unconstrained this can lead to an avoidance action, e.g.running away.However, alternative action predictions might result from supervised forward modelling.For example, the prediction of an elated reaction and accepted date might allow one to overcome the threat of anticipated rejection and ask your secretly beloved out for a date.This forward modelling is a re-prediction of action consequences and might lead to a more favorable predicted outcome and an alternative behavior.We argue that this mechanism can also account for more abstract emotion regulation commonly referred to as 'reappraisal of emotional situations'.For instance, post-hoc imagination of your beloved being elated by your request may elicit the same felt tendency that would have been elicited had the situation been factual, i.e. to actually approach your beloved, with associated psychophysiological changes and subjective feeling of happiness.

Comparative argument: the brain (and emotion) developed to guide survival-relevant action
Phylogenetically, neural systems build on negative feedback control circuits that adapt behavior when resources run low or when a threat is perceived (Branco and Redgrave, 2020;Cisek, 2019;Feinberg and Mallatt, 2016).Extension of brain circuits in hierarchical feedback loops allows action-selection based on more abstract control parameters (Cisek, 2022(Cisek, , 2019) ) such as predicted future homeostatic deviations, also called allostatic control (Katsumi et al., 2022;Schulkin and Sterling, 2019).Emotions are no exception to this organization (Barrett, 2017;Ledoux and Daw, 2018;Ledoux, 2007).They are elicited with a purpose, an impetus (Ridderinkhof, 2017); to promote those behaviors that benefit preservation of the organism and to drive the organism away from those situations that are likely to end up causing harmthey are pre-motor faculties in the sense that they inform which action should be selected to deal with relevant current or anticipated circumstances (Barrett, 2017;Fine and Hayden, 2022).To put those into context, consider William James' famous example of encountering a bear in the forest.The observation that there is a bear that might eat you will result in a strong action tendency to run away in order to increase the distance to the threat (valence).This action tendency is accompanied by corollary discharge through physiological and visceral prediction systems that bring about the changes allowing you to do so effectively (arousal), all elicited in order to minimize the possibility of being eaten.Negative feedback control refers to the aim of the elicited (flight) behavior, minimizing the chance of being eaten by increasing the distance between you and the bear.This is achieved most effectively through taking action (Branco and Redgrave, 2020;Cisek, 2019).The subjective component (e.g. the feeling of fear) that accompanies these encounters might reflect the systems estimate of how well the current strategy is performing in achieving its goal (Bach and Dayan, 2017), in other words the difference between the current and the desired state.This estimate is based on the combined interpretation of physiological-, visceral-and sensory signals (Petzschner et al., 2021;Sennesh et al., 2021;Seth and Tsakiris, 2018) and results from the inference made on the cause of the change in sensory-, visceral and physiological signals.When running away is successful in increasing the distance, the subjective fear reduces, along with the predicted visceral and physiological needs, thereby Fig. 2. Key figure: Forward modelling in emotional action-selection and regulation.Panel A, left side: To respond to allostatic (e.g.predicted hunger) or acute homeostatic challenges (e.g encountering your secret beloved) multiple potential actions are specified on the fly in parietal-and (pre-)motor circuits, based on currently available affordances (purple lines) partly derived from direct inputs from sensory cortices (shaded lines).For instance, "ask for a date" (i.e.approach) versus "run away" (i.e.avoid; Fig. 1).Potential outcomes of these action plans are predicted (blue) based on learned or innate action-outcome predictions ("scoring a date" versus "being laughed at and embarrassed").Corollary discharge also routes these action plans through systems involved in physiological and visceral predictions that prepare the body for rapid and potentially vigorous responding.In the case of unsupervised actionoutcome prediction, Pavlovian stimulus-action associations may lead to biases in action selection.We think this might coincide with what emotion researchers call appraisal and automatic action tendencies.During ongoing action, predicted action outcomes, as well as physiological and visceral requirements of the unfolding motor plan are continuously monitored and updated through these same loops, whilst potential benefits of the best alternative action are monitored in lateral prefrontalparietal circuits (purple lines in panel B).In this framework, emotional-action control becomes an instance of action selection.Arbitrating between Pavlovian and instrumental actions, but also between multiple potential instrumental action strategies, depends on competing outcome predictions, which in some cases (e.g.overriding your nerves to engage in conversation) might require a supervised re-prediction of prepotent predicted outcomes (e.g.telling yourself 'Maybe they like me back').Panel B: Emotional action-selection and abstract (counterfactual) emotion regulation (e.g.reappraisal) might depend on the same action-oriented forward and inverse modelling, without overt motor output.Say that you have asked your crush on a date and they laughed at you.Reliving this situation re-activates the executed action plan without actual sensory input (Buzsáki et al., 2014), and engages the circuitry for predicting its outcome.Reliving this situation can elicit similar feelings of shame, including physiological and visceral responses.Regulating this emotion can then be interpreted as supervised re-prediction of alternative action outcomes, "they laughed because I was funny, not because I am ridiculous".If this re-afferent prediction is sufficiently probable, physiological and visceral responses change through corollary discharge.In contrast, in cases where sensory information conflicts too much with the re-afferent prediction one might discard the alternative interpretation, in which case reappraisal fails.Amygdala (Amy), ventro-medial prefrontal cortex (vmPFC), dorsal-and pregenual anterior cingulate cortex (dACC, pACC), hypothalamus (Hypo), periaqueductal grey (PAG).For simplicity we omitted a.o.loops through the ventral pathway, hippocampus and striatal/BG circuits.Figure is inspired by (Barrett and Simmons, 2015;Chanes and Barrett, 2016;Cisek, 2007;Cisek and Kalaska, 2010).Shading around lines is applied to emphasize changes between panel A and B.
B. Bramson et al. essentially "regulating" the emotion.In absence of the bear, those same increases in respiration and heart-rate can be interpreted as a consequence of physical effort rather than fear.Although this example focusses on reactive instances of emotion, one can also think of allostatic examples, for instance the anxiety that is elicited at the potentiality of encountering a bear, which might bias behavior away from exploring deeper parts of a forest.This line of argumentation naturally leads to a prominent role of action in emotion and its regulation, implemented by pre-motor systems responsible for action-selection (Fine and Hayden, 2022;Pine et al., 2021).
It is easy to see how externally encountered aversive states can be minimized through action.Similar arguments can be made for appetitive motivational states.Examples include the desire for food, solved through ingestion and consequent interoceptive feedback from gustatory receptors (Critchley and Harrison, 2013).Other examples are sexual desires, effectively "regulated" through action; or more abstract states such as curiosity, pride or pleasure when envisioning a scientific career, which can temporally be satisfied (i.e.regulated) by publishing a good paper.
As the latter examples illustrate, human emotional states and behaviors can have a more complex structure than the acute Pavlovian flight-fight decisions sketched in the example of the bear encounter (Ramstead et al., 2016;Whiten et al., 2017), and emotional states are no longer necessarily dependent solely on acute internal or external events.Those states can be counterfactual or prospective (Corcoran et al., 2020;Pezzulo and Castelfranchi, 2009;Pine et al., 2021), however, it is unclear whether -from a theoretical perspective -action-selection can still be the regulator when emotional states are based on abstract or counterfactual inference.

Theoretical argument
Recent work has reframed several instances of human cognition in the context of action-prediction, detached from overt motor output (Barsalou, 2008;Buzsáki et al., 2014;Corcoran et al., 2020;Pezzulo, 2011;Pezzulo and Cisek, 2016;Winkielman et al., 2018).We have seen that structurally simple neural systems consist of (relatively) direct situation-action loops where approach or avoidance are direct reactions to internal or environmental state changes (Cisek, 2019;Ledoux and Daw, 2018).Crucially, the development of multiple neural systems on top of basic sensorimotor loops allows activity to be maintained and detached from overt motor output (Buzsáki et al., 2014;György Buzsáki, 2019).The ability to sustain activity without direct dependency from input and output allows for simulation of multiple action strategies to allostatically predict their potential outcomes in absence of actual movement, an example of forward modelling (Box 1; (Franklin and Wolpert, 2011;Gallivan et al., 2018;Pickering and Clark, 2014;Ridderinkhof, 2017;Schubotz, 2007;Wolpert and Kawato, 1998)) .
During overt motor-control, forward models allow for the continuous comparison of incoming sensory information to predicted outcome effects, yielding prediction errors that can serve to optimize behavior (Box 1).However, prediction mechanisms can also be used in 'simulation mode' only, for instance by inhibition of motor output, or through reducing precision on sensorimotor predictions (Adams et al., 2013;Friston, 2011).This 'mental simulation' which predicts the consequences of hypothetical behavioral strategies, has been suggested to underlie the ability to covertly compare potential actions and predict their hypothesized outcomes (Barrett, 2017;Barsalou, 2008).Here, we integrate this view into the context of emotion regulation.We propose that acute emotional action control, but also instances of more abstract emotion regulation (such as cognitive reappraisal) might be based on similar forward and inverse modelling of hypothesized actions and their outcomes.Concretely, the abstract regulation of emotional responses requires situational re-predictionsupervised forward modellingthat provides an alternative to the automatic emotional response.This process can be decoupled from acute motor output, but still grounded in those same pre-motor systems developed for selecting actions on the fly (Fig. 2; panel B).
The notion that emotion regulation follows principles of sensorymotor control would predict that emotion processing and regulation would activate neural motor circuits.
Box 1 forward modelling to anticipate action-effects.
Forward modelling has been most clearly described for instances of sensorimotor control (McNamee and Wolpert, 2019).When performing an action, sensory feedback about the consequences of that action will always lag behind its execution, and will therefore always arrive too late to trigger effective corrections to the motor plan.In order to allow for rapid and dynamic adaptation of actions, motor commands are accompanied by visceromotor predictions that recruit allostatic bodily responses, and efference copies (more constrained version of the corollary discharge) to lower-level sensory cortices (McNamee and Wolpert, 2019).These efference copies are exact copies of the motor plan that can be used to simulate the sensory consequences of those actions; a process called forward modelling.This way, potential disturbances in the action trajectory can be compensated before they interfere with the action.Efference copies also allow the subtraction of self-generated sensory input from the total sensorium to allow a dissociation between self and other (this is why you cannot tickle yourself).
The outcome of these forward models can be used to inform the action that is to be taken, or the adaptation of ongoing behavior that is required.This can then be communicated with multiple other systems through corollary discharge, also called spreading activation (Crapse and Sommer, 2008).In the context of emotional responses, corollary discharges are important for mobilizing physiological responses in homeostatically challenging situations.For instance, when confronted with a threat (e.g. a bear), the forward modelling of the consequences of the action plan to flee might engage physiological responses that allow for a vigorous response through corollary discharge to systems involved in autonomic arousal.The knowledge of better coping strategies, e.g.standing your ground and making loud noises, or approaching the bear to assert dominance require alternative action plans that can be selected based on learned outcome predictions (Etkin et al., 2015;Soltani and Koechlin, 2022), which in turn change physiological responses again through corollary discharge.
Crucial for our point, forward models can also be run 'offline' in anticipatory or counterfactual processingprediction -of ongoing or hypothesized actions and their consequences (Pezzulo et al., 2015;Ridderinkhof, 2017) a process also called mental simulation (Barsalou, 2008).In other contexts, forward models can be invertedinverse modellingto infer the goals and emotions from behavior and movements of others (Barsalou, 2008;Bastiaansen et al., 2009;Gallivan et al., 2018;Schubotz, 2007;Wolpert and Flanagan, 2001).Below we will argue that this organization can provide a way to understand both emotional-action selection and cognitive emotion regulation through forward modelling of action outcomes (real or counterfactual).Forward models can influence emotional states through corollary discharge, even in the absence of overt movement.
B. Bramson et al.

Empirical argument: emotion processing and regulation consistently engages neural motor circuitry
As illustrated in the introduction, appraisal of (internal and external) states relevant to the organisms survival automatically elicits action tendencies based on the available affordances (Moors et al., 2019;Ridderinkhof, 2017).This action-readiness is mirrored in experimental findings consistently reporting activity in motor regions in response to potential homeo-or allostatic challenges.For example, the presentation of affect-laden stimuli (e.g.threatening faces) elicits increased activity in primary and supplementary motor cortices (Bramson et al., 2018;Engelen et al., 2018;Hermans et al., 2013;Meaux et al., 2019;Van Den Stock et al., 2011), and in sensorimotor nuclei such as the periaqueductal gray (PAG) and hypothalamus, involved in organizing defensive responses to survival-relevant information (Hermans et al., 2013;Mobbs et al., 2007;Qi et al., 2018;Silva and Mcnaughton, 2019).Other studies have shown that the presentation of affective information enhances excitability in cortical motor areas, a situation that might reflect preparation for taking rapid action to control the emotional state (Frijda et al., 2014;Hajcak et al., 2007;Van Loon et al., 2010).Finally, although most studies that employ threat of shocks do not study action preparation and motor-cortical activation directly, several of those studies have shown that the magnitude freezing induced by threat-of-shock is linked to faster subsequent reactions, without cost to accuracy, suggesting action preparation (Hashemi et al., 2019;Klaassen et al., 2021) The influence of freeze on action preparation is likely mediated through midbrain influences on (pre) motor circuits (Qi et al., 2018).
In addition to preparing the body for taking action, motor cortical activity is also consistently elicited during other instances of emotional processing.For instance, fearful faces in others are recognized best if one is in a bodily state that mimics the manifestations of fear (Pezzulo et al., 2018), whereas mimicry of emotional expressions elicits those feelings, potentially achieved by covertly simulating the neural activity needed to create those actual emotions (Palagi et al., 2020;Wood et al., 2016).The hypothesis that covertly simulating action through motor-cortical activity is necessary for understanding emotion is supported by findings that motor cortices contribute strongly to classification of different emotions (Diano et al., 2017;Saarimäki et al., 2016), and by studies showing that the ability to recognize or simulate emotions is diminished when activity in motor areas is disrupted by TMS (Davis et al., 2017;Keysers et al., 2018;Oliveri et al., 2003;Rochas et al., 2013), impossible due to tissue damage (Adolphs et al., 2000;Keysers et al., 2018), or involved in overt motor tasks (Davis et al., 2017).
How about emotion regulation?It is likely that the motor cortical activity observed during emotional situations reflects the impetus for action that comes with the survival-relevant nature of emotion-eliciting situations (Powers and LaBar, 2019;Ridderinkhof, 2017).Strikingly, during emotion-regulation the activity in primary-and pre-motor cortices increases even further, over and above the increase already seen with passive emotional picture viewing (Goldin et al., 2008;Kohn et al., 2014;Morawetz et al., 2020Morawetz et al., , 2017)).For instance, reappraisal of emotional stimuli as compared to passive viewing often elicits stronger activity in regions such as SMA (Kohn et al., 2014;Morawetz et al., 2020;Powers and LaBar, 2019), premotor and primary motor cortex (Buhle et al., 2014;Goldin et al., 2008), as well as anterior prefrontal cortex also involved in emotional-action control (Bramson et al., 2020b;Koch et al., 2018;Morawetz et al., 2017).Moreover, these motor-cortical activity patterns often show as strong or stronger correlations with emotion regulation effectiveness than more rostral frontal regions or the amygdala (Ochsner et al., 2002) and show changes during real-time feedback emotion regulation training (Herwig et al., 2019).It seems parsimonious to interpret these correlational observations as an indication that both emotion processing and control are supported by sensorimotor or action-selection processes.Making those influences explicit, also in abstract emotion regulation instances, can provide additional insight into the mechanisms and limitations of emotion regulation strategies.However, it does raise the issue of qualifying the computations implemented in the extended (pre)motor system during emotion regulation.

Emotion regulation as action control, detached from motor output
The comparative, theoretical, and empirical arguments presented above open the way to an emotion regulation theory that emphasizes action control mechanisms.More precisely, emotion regulation can be conceptualized as a continuum, ranging from selection of action Box 2 Emotion augmentation through simulation is common and useful in everyday-and clinical situations.
The supervised (re-)prediction ("simulation") of emotional actions and situations in response to internal or external cues is apparent in daily life and clinical instances of emotion regulation.By way of example: you may have found yourself fantasizing about kissing the target of your affection, without overtly acting on this fantasy.On the other hand, imagining how situations could have been handled differently after an embarrassing mistake, and prospectively simulating how situations might pan out can increase coping with these situations in the future (Suddendorf and Corballis, 2007).
Several therapeutic treatments utilize similar mental simulation to treat emotional disorders.Imaginary exposure therapy, in patients who developed post-traumatic stress-disorder after an assault during which they were unable to move, invites patients to imagine an alternative strategy, for example to stand up and stop their assailant.Those techniques have been found to be effective not only for PTSD but also other conditions characterized by intrusive cognitions (Holmes et al., 2021), for example, in fear of spiders.Before treatment, the imagination of a spider without actual sensory input led to significant increases in subjective fear, skin conductance and activity in neural fear circuits, including amygdala and ACC.A ten minutes session of imaginary exposure therapy can significantly reduce those fear-related processes (Hoppe et al., 2021).
Moreover, in a recent large-scale prospective study (Kaldewaij et al., 2021), we have shown that neural activity elicited during acute emotional action control can predict clinical outcomes, and more strongly so than self-report and behavioral measures.Here, police recruits newly admitted to the academy and at high risk for trauma-exposure, were followed throughout their training, involving multiple exposures to potentially traumatic events.The development of PTSD symptoms after this period of trauma exposure could be predicted based on baseline emotional-action control-related activity in lateral prefrontal cortex.Officers that showed more action-control related activation at baseline were more resilient against the adverse consequences of trauma.These acute action-control strategies provide cues for potential action-oriented interventions aimed at increasing resilience.For instance, the same emotional action-control ability that is predictive of trauma-resilience can be improved by synchronizing prefrontal and sensorimotor cortex using non-invasive brain stimulation (Bramson et al., 2020).
possibilities alternative to Pavlovian tendencies, to abstract reprediction of alternative hypothesized actions.We suggest both are based on forward modelling, grounded in the (pre-) motor system responsible for action-selection (Barsalou, 2008;Fine and Hayden, 2022;Mobbs et al., 2015;Pezzulo and Cisek, 2016;Winkielman et al., 2018).When confronted with a challenging emotional situation, unsupervised forward modelling can influence action selection towards efficient coping, based on predictions resulting from innate or learned stimulus-action mappings (Bach and Dayan, 2017;Huys et al., 2012;Mobbs et al., 2020) -Pavlovian action tendencies (Fig. 1; (Bach and Dayan, 2017)).When contextual constraints cause a conflict between Pavlovian predictions and current goals, alternative actions can be modelled in higher levels of the pre-motor hierarchy (Badre and Nee, 2018;Fine and Hayden, 2022;Mansouri et al., 2017;Pezzulo and Cisek, 2016), for instance based on strategies acquired through reinforcement learning (Etkin et al., 2015), and selected based on their predicted outcomes and estimated reliability.In cases where alternative strategies are deemed unreliable, new action strategies can be explored through trial and error, or by combining parts of existing or observed action strategies stored in long-term memory into novel strategies (Soltani and Koechlin, 2022).
During model-based emotion regulation, for instance when reliving an emotional event, the same combination of memorized and newly explored strategies can result in mental simulation that changes the original memory in a way that has re-afferent influences on visceral, physiological and subjective processes.Take the example of being laughed at by your crush upon asking them for a date.Reliving this situation might entail a covert replay of the neural activation elicited during the actual situation (Foster, 2017;Liu et al., 2021), meaning that the executed action plan is re-routed through circuitry involved in action-outcome predictions leading to the same conclusion "they laugh and will wave me off afterwards", and the same visceral, physiological (e.g.blushing, sweating) as well as subjective (e.g.feeling shame) responses.A subtle re-prediction of this situation however, allows the instigation of alternative interpretations, "the laugh will be followed by a hand reach because I was funny"; Fig. 1.Critical for the field of emotion regulation, the corollary discharges that accompany those re-afferent sensorimotor predictions can account for what is often called 'cognitive reappraisal', that is a reinterpretation or mental rewriting of the event by combining the experience with alternative simulated outcomes.This can occur in the absence of overt motor output, yet in the presence of full blown changes in psychophysiological, visceral, and subjective components because the anticipated outcomes send corollary discharges to areas involved in visceral and physiological responses.

Neural implementation
To see how abstract emotion regulation can depend on action circuitry that initially developed for dealing with events with acute homeostatic relevance, we first sketch the circuitry implemented in overt emotional action control.
Encountering a situation with acute homeostatic relevance, such as the detection of a threat, elicits multiple potential action strategies in (pre-)motor and parietal systems, that are computed on the fly based on available affordances (Badre and Nee, 2018;Cisek, 2022;Feinberg and Mallatt, 2016;Frijda, 1986) (Fig. 2).These strategies compete for execution based on their predicted outcomes, computed by routing corollary discharges through circuitry involved in action-outcome estimations such as hippocampus, basolateral amygdala, striatum, medialand orbitofrontal cortices [20,81− 85].Additional corollary discharges are routed through systems such as the PAG, medulla, hypothalamus, central amygdala and pregenual ACC; areas involved in prediction (and eliciting) of upcoming autonomic and physiological requirements that are necessary to cope with the situation (e.g.increased blood flow to muscles and decrease of blood flow to digestive organs (Kozlowska et al., 2015)).Whereas prospective action simulation in safe situations can provide strategies for coping (Mobbs et al., 2020;Pine et al., 2021), growing response urgency due to increased imminence might short-circuit competition by lowering decision thresholds (better safe than sorry) or by employing faster, pre-programmed or overlearned Pavlovian responses (Bach and Dayan, 2017;Mobbs et al., 2020).Crucially, however, even the selection of rapid, pre-potent action strategies can depend on forward modelling that takes into account current environmental circumstances (Dayan and Berridge, 2014), as is evidenced by the observation that even time-constrained flee-decisions are modulated by current environmental affordances such as the availability of shelter (Branco and Redgrave, 2020).The action strategy yielding (at present) the most beneficial predicted outcome is selected for execution (note that this can also be a non-move strategy when evidence is unclear, or a return to normal functioning if a situation is considered non-threatening; Mobbs et al., 2020Mobbs et al., , 2015)).Continuous feedback loops provide updated predictions during the unfolding of the selected action strategy to maintain the behavior until the trigger state has dissipated (negative feedback control; Cisek, 2019) or until counterfactual strategies become more beneficial (Koch et al., 2018).What we call emotion is the inference that is made on the whole process of forward modelling of action plans, including its (continuously predicted) allostatic, physiological and autonomous consequences (Bach and Dayan, 2017;Barrett, 2017;Barrett and Westlin, 2021;Barsalou, 2017;Feldman Barrett and Finlay, 2018;Frijda, 1986;Seth, 2015Seth, , 2013)).
This framework can naturally be extended to instances of emotional action-control in which contextual or instrumental demands conflict with a prepotent emotional action tendency.This can require arbitration between Pavlovian-and rule-based or instrumental behavior.Emotional action control then becomes an action selection featwhere the selected action strategy is based on predicted outcomes of Pavlovian versus alternative (e.g.rule-based) strategies that might have been acquired through various forms of learning such as reinforcement (Etkin et al., 2015), learning by observation, or cultural transmission (Whiten et al., 2017), or on combining previous knowledge into counterfactual predictions (Soltani and Koechlin, 2022).There is evidence that in this type of emotional action control counterfactual strategies are continuously monitored in rostro-lateral prefrontal-parietal circuits to rapidly implement alternative actions should the current strategy become unreliable (Boschin et al., 2015;Bramson et al., 2020a;Koch et al., 2018;Mansouri et al., 2017;Miyamoto et al., 2021;Volman et al., 2011).Crucially, mechanistically informed brain stimulation targeting these very same prefrontal action-selection mechanisms can improve the individuals' control over emotional behavior (Bramson et al., 2020a).
Central in our view is the notion that emotion regulation strategies traditionally seen as cognitive or verbal, actually depend on the same action-prediction computations, detached from acute motor output.Take the example of the reappraisal of an emotional memory, say being laughed at when you asked your crush on a date.Re-living this situation involves re-activating the executed action plans (Buzsáki et al., 2014;Schubotz, 2007), along with a re-prediction of the consequences (in this case based on outcome memory).Routing this action plan through the same circuitry that was involved during the actual situation can recreate those feelings, including physiological and visceral changes.During reappraisal, alternative actions and outcomes are modelled that yield alternative predicted consequences in sensorimotor cortical areas, as summarized above, in addition to changes in agranular limbic systems that predict visceromotor responses to these modeled events to prepare the body for action (Barrett and Simmons, 2015;Chanes and Barrett, 2016).These predictions can lead to changes in bodily sensations through corollary discharge, in addition to changing physiological responses and subjective components to the extent that those simulations are in line with the desired goals (Fig. 2, panel B; in red).

Concluding remarks
This opinion piece aims to bring together work on emotion B. Bramson et al. regulation and emotional action control under a single conceptual framework; action-outcome prediction.Conceptualizing emotion regulation as such is a direct continuation of recent trends emphasizing the importance of phylogenetic plausibility to constrain thinking on brain functioning and cognition (Buzsáki et al., 2014;Cisek, 2019;Pessoa et al., 2022;Pezzulo and Cisek, 2016) and predictive processing accounts of emotion and action (Barrett, 2017;Schubotz, 2007;Seth, 2013;Seth and Friston, 2016).Our hope is that conceptualizing emotion regulation in this framework can accommodate the large extent of work on cognitive emotion regulation strategies (Etkin et al., 2015) and still allow for acute emotional action control (Bach and Dayan, 2017;Mobbs et al., 2020), whilst providing explicit and plausible neural mechanisms that implement both acute and more abstract forms of emotion regulation.
Our framework leans on earlier cybernetic and action-theoretical frameworks of emotion and its regulation (Frijda, 1986;Gross, 2015a;Mesquita and Frijda, 2011;Tamir, 2021).However, as compared to the former, we frame emotion itself as a control-state that is elicited to achieve a behavioral goal (Blakemore and Vuilleumier, 2017;Ridderinkhof, 2017), rather than the change in emotional state being the goal in itself (Tamir, 2021).Pavlovian action tendencies might then regulate these states by providing innate or learned action strategies that have proven beneficial in the past.During the developing action, subjective feelings can reflect the estimate of how well in the agent performs in regulating this state (Bach and Dayan, 2017).Although sometimes these Pavlovian action strategies are sufficient, in other cases we need to rely on more abstractmodel basedstrategies, for instance acquired through reinforcement learning (Etkin et al., 2015), cultural transmission (Whiten et al., 2017), or the exploration of novel strategies (Koch et al., 2018;Soltani and Koechlin, 2022).Our main contribution here stems from the suggestion that these more abstract model-based forms of emotion regulation might depend on the same mechanisms selected through evolution because of the action-selection benefits they provided (Cisek, 2022).Namely, the ability to envision alternative action strategies to allow behavior to diversify and adapt to continuously changing environmental demands that come with human cooperative society (Hare, 2017;Whiten et al., 2017).Crucially, this way of formulating allows the specification of concrete neural implementations derived from action-selection literature to inform potential emotion-control implementations, such as theta-gamma coupling between prefrontal and sensorimotor cortices (Bramson et al., 2018;Voytek et al., 2015;Voytek and Knight, 2015).
Interpreting emotion regulation as an action-control continuum has several advantages.First, it provides an integrated mechanistic theory that accounts for different types of emotion regulation that are traditionally treated in separate literatures, with as extreme examples cognitive reappraisal versus emotional action control.Second, the account is grounded in well-established biological, neurocognitive and feed-forward and feedback models (Barrett and Simmons, 2015;Schubotz, 2007;Seth, 2013).Therewith, it does acknowledge subjective and appraisal components of emotion, albeit with more precise neuro-computational predictions than current cognitive emotion regulation theories.Third, formulating emotion regulation in control-theoretic terms (Barsalou, 2008;Buzsáki et al., 2014;Pine et al., 2021) provides a plausible phylogenetic perspective on the potential development of emotion-and emotion regulation capacities, with terminology that multiple sides of the affective neuroscience field might accept (Mobbs et al., 2019), easing inter-species comparisons.Fourth, it provides clear and proven starting points for improved symptom prediction and treatment of affective disorders (Bramson et al., 2020a;Kaldewaij et al., 2021).
The validity of our framework depends on several assumptions.First, we framed our work on a solid basis that suggests that our understanding of the world and the actions we can take in that world is grounded in our motor system (Barsalou, 2008;Buzsáki et al., 2014;Cisek, 2019;Pezzulo, 2011;Schubotz, 2007).Second, we make the assumption that an emotional state is elicited to signal and cope with the discrepancy between current or anticipated states and the goal states of an organism.Although this might lead to multiple unsupervised strategies, the extent of the deviation from the internal set-point, combined with the predicted outcome might determine which strategy is selected.
There are still many outstanding questions, some important ones are summarized below.For instance, in our current framework we did not specify how higher-order appraisals such as the sense of agency or attributed intentions can be explained in terms of sensorimotor predictions.Nor do we explain all instances emotion regulation.For example, it remains to be seen whether self-oriented reappraisal such as distancing or taking third-person perspectives can also be explained through supervised action-simulation.Although we do not consider it unlikely that our theory can be extended to also encompass those instances, we hope that future work will clarify the boundaries of the current action-oriented approach.Nevertheless, we think that to address those issues, we should attempt to avoid further separation between cognitive and action-based emotion regulation strategies.Such subdivision seems to have created an intellectual dead end and is inherently counterproductive, creating a false dichotomy between regulation strategies that might share a mechanistic basis.We hope to have laid the grounds for an exploration of common denominators of variations in emotional regulation by focusing on shared underlying circuitry and development.

Outstanding questions
1.Although similar neural circuits have been established for acute and abstract emotion regulation (Koch et al., 2018) the question remains whether strategies used to improve acute emotional-action control, such as neural stimulation (Bramson et al., 2020a) can be adapted to also improve more abstract emotion regulation such as reappraisal.This could potentially be a breakthrough for clinical practice where there is ample room for boosting treatment effects.2. Relatedly, it should be investigated whether cognitive reappraisal techniques can be improved by concretizing instructions, e.g.towards re-simulating actions and their outcomes.This might enable easier translation of experimental findings towards clinical applications (Kredlow et al., 2022).3. It is well established that efficacy of emotion-regulation strategies such as reappraisal is context dependent: distraction works better in high-intensity contexts, and re-appraisal works better in less-intense emotion contexts (Sheppes et al., 2014).Future work could test how to exploit that difference to optimize the efficacy of regulation strategies.For instance, it might be beneficial to emphasize overt avoidance in high-intensity situations, and mental prediction of alternative course of action in low-intensity contexts.4. Can an action-oriented view of emotion regulation be extended to more higher-order appraisals such as the attribution of agency or intentionality when encountering a perceived situation that elicits an emotional response.It is plausible that similar simulations of others actions can be employed to inverse-model another's intentions, as is the case for less abstract inferences such as another's pain (Gallo et al., 2018). 5. Given that emotional expressions and imagery can be decoded most accurately from sensorimotor cortices, it becomes interesting to test whether changing emotional content through e.g.reappraisal changes the content that can be decoded from sensorimotor areas.6.Although there is some TMS work on model-based decision making (Smittenaar et al., 2013), many of the neural systems involved in Pavlovian control, such as the PAG and amygdala, were outside of the range of noninvasive neuromodulation until recently.The translation of focused ultrasound for neuromodulation (Folloni et al., 2019)  their influence on different phases of emotional action selection (Mobbs et al., 2020).

Glossary
Affordances: The set of action-possibilities of an agent provided by the current environment.Allostasis: The predictive maintenance of the internal milieu of a system within a set of internally specified bounds.Appraisal: The process in which a situation is rapidly assessed in terms of its relevance for the organism.That is, determining whether a state of affairs is potentially good for you or bad for you, whether it needs acting upon given the current affordances.In other words, the recognition of current affordances and how they are potentially beneficial or detrimental for the goals of the organism.Corollary discharge: Originally defined as copies of motor commands transmitted to multiple levels of the sensory hierarchy to anticipate ones' own influence on sensory changes.Recently, corollary discharge has been suggested to also encompass the signaling of more abstract action plans to sensory and predictive systems throughout the brain (Crapse and Sommer, 2008;Subramanian et al., 2019).These signals can be used to predict potential visceral and physiological requirements for action, and the expected consequences of those actions in terms of sensory feedback because they allow the simulation of action outcomes.In our framework, the immediate evolving physiological and visceral responses in affectively-laden situations suggests that appraisal of affordances, and resulting action tendencies, might provide corollary discharge to rapidly prepare the body for this action.Crucial for our framework is the notion that the extension of motor circuitry that allows action simulation uses the same corollary discharge circuitry.Therefore, also simulated actions can result in changes in physiological, visceral and subjective responses.Emotion: Emotions are states that combine psychophysiological and visceral responses with subjective and behavioral components such as action tendencies.In line with interoceptive inference theories (e.g.Seth, 2013) we view emotional states as resulting from inferences over the causes of intero-and exteroceptive signals (appraisal).These signals are interpreted (appraised) in terms of their relevance and required behavioral strategy, where the subjective feeling component reflects the discrepancy between the current or anticipated state and the homeostatic or allostatic goal.Visceral predictions and physiological responses are elicited to be able to maintain actions until the difference between the animals state and the internal set-point is minimized.Emotion regulation: The process of changing subjective, physiological and visceral, or behavioral features of the emotion, resulting in an altered mental state.Here, we advocate that emotion-regulation can be seen as a form of action selection.When there is a discrepancy between the current or anticipated state of the organism and its intended goal state, an emotion is elicited.Performing an action to reduce that discrepancy regulates the emotion by reducing this discrepancy.This consequently changes the subjective, physiological and visceral responses.Extended pre-motor system: The dramatic extension of prefrontal systems in the primate lineage can be interpreted as extension of neural circuits involved in setting the stage for appropriate action ('pre-motor') (Fine and Hayden, 2022).This definition builds on recent work suggesting that abstract 'cognitive' capacities such as memory, attention and decision making have developed as hierarchical extensions of action-selection circuits (Cisek, 2019).Forward modelling: The prediction of future states of the system, achieved by modelling the expected consequences of an unfolding situation.In the context of actionselection, one can envision simulating the potential consequences of an action (one's own or another's) in terms of sensory changes or changes in the state of the world.Forward modelling can happen relatively automatically based on learned, yet implicit, associations (unsupervised forward modelling); or more explicitly and counterfactually, involving alternative actions that are not based on implicit stimulusoutcome associations (supervised forward modelling).Appraisal might be seen as a rudimentary form of forward modelling, determining what the relevance of a certain situation is, and whether a response is warranted.Homeostasis: The reactive maintenance of the internal milieu of a system within a set of specified bounds.Negative feedback control: A concept of cybernetic theory.Negative feedback control refers to the ability of a system (biological or otherwise) to minimize deviations from a set-point.This action brings the system closer to its desired state, thereby resolving the need to act.The classical example is a thermostat, that increases or decreases heating to maintain the room temperature to a certain temperature.Biological systems are also often described in terms of negative feedback control, for instance changing behavior to maintain nutritional balance.Importantly, negative feedback control can be extended through prediction of upcoming deviations from internal set-points, allowing for allostatic behavior.Model based control: a form of control that is based on internal representations of the environment (internal models) and that can be used to simulate potential outcomes of action strategies depending on the context.Pavlovian control: innate or learned action-selection algorithms that dictate preprogrammed responses to certain situations (e.g.action invigoration when confronted with potential rewards; behavioral inhibition in response to punishment).These action tendencies are thought to be formed in stable environments where the same stimulus-action association reliably leads to beneficial outcomes (Bach, 2015), which allows shortcuts in computation by association global action strategies with outcomes to achieve rapid and accurate responding (model-free behaviors).Although these tendencies might be beneficial overall in situations where reward is not contingent on action-strategy, they are relatively inflexible and less adaptable in situations in which reward can be controlled by taking strategic actions (Dorfman and Gershman, 2019).Phylogenetic: The study of evolutionary adaptations that occurred in a species-specific evolutionary history.Reappraisal: A specific emotion-regulation strategy often implemented in emotionregulation studies.The strategy consists of a re-interpretation of emotional situations/inferences.We argue here that this strategy can be understood in terms of action and action-outcome simulation.
towards human applications might create the possibility to test Pavlovian versus model-based emotion regulation influences and B. Bramson et al.