What is play in predictive minds: A mechanism of prediction error minimization, whereby the brain attempts to reduce the mismatch between how it predicts the world to be and how the world actually is

Play in Predictive Minds: A Cognitive Theory of Play. Marc Malmdorf Andersen email the author, Julian Kiverstein, Mark Miller, Andreas Roepstorff. Psychological Review. jun 2022. https://doi.org/10.1037/rev0000369

Abstract: In this article, we argue that a predictive processing framework (PP) may provide elements for a proximate model of play in children and adults. We propose that play is a behavior in which the agent, in contexts of freedom from the demands of certain competing cognitive systems, deliberately seeks out or creates surprising situations that gravitate toward sweet-spots of relative complexity with the goal of resolving surprise. We further propose that play is experientially associated with a feel-good quality because the agent is reducing significant levels of prediction error (i.e., surprise) faster than expected. We argue that this framework can unify a range of well-established findings in play and developmental research that highlights the role of play in learning, and that casts children as Bayesian learners. The theory integrates the role of positive valence in play (i.e., explaining why play is fun); and what it is to be in a playful mood. Central to the account is the idea that playful agents may create and establish an environment tailored to the generation and further resolution of surprise and uncertainty. Play emerges here as a variety of niche construction where the organism modulates its physical and social environment in order to maximize the productive potential of surprise.

Keywords: play, learning, predictive processing, surprise, niche construction

Discussion

If play is, at its core, the deliberate seeking and creation of surprising situations, this has important implications for learning, niche construction; for current understandings of playfulness as a general mood state; as well as methodological implications for future research on play in humans.

Play, Learning, and Niche Construction

Humans in general, and children in particular, play not only to chase slopes of error reduction but also to actively build and create such slopes of error reduction. This perspective may be relevant for recent work in evolutionary biology that addresses predictive processing and niche construction. Predictive processing extends to nonhuman animals as well because prediction error minimization is believed to be a universal biological process in which organisms attempt to keep themselves within expected sensory and physiological states given their species-specific prestructuring and the niche they inhabit (Friston, 2010). In evolutionary biology, niche construction refers to the process of organisms modifying their environment, thereby steering their own and others’ evolutionary trajectory (Laland et al., 2015). Recent arguments suggest that the mathematics of predictive processing can be used to model the effect of niche construction on biological evolutionary processes (Constant et al., 2018).

From this perspective, niche construction is a way for organisms to efficiently minimize prediction error by manipulating the environment to conform to their own expected states. Thus, an organism’s species-specific prestructuring may prompt it to build a nest or a burrow, ensuring that expectations about things such as wind speed or temperature are effectively met. Niche construction may therefore be seen as a form of active inference, where the organism manipulates the environment to fit its own expectations. In many cases, animals are born into an already altered environment fit to suit their species-specific prestructured expectations, for example, ants in an anthill; beavers in a lodge; or humans in a house (Constant et al., 2018). What this account may have overlooked, however, is that a handful of species, notably the most intelligent, regularly engage in playful behavior after their basic expectations have been met. Human children actively seek and create situations that they expect to be surprising in an effort to reduce uncertainty. When the environment offers no uncertainty, children will readily modulate it in such a way that it becomes error-inducing.

It is an open question as to whether this may also be the case for certain nonhuman species famous for their playful inclinations. Dolphins, for instance, can often be seen creating bubble rings by exhaling air through their blowholes, which they subsequently play with in a variety of ways. Some dolphins have even been observed to produce multiple rings that they then join together, or push one through another (Janik, 2015). Similarly, several populations of Bornean orangutans have been documented building nests for social play, and object-substitution and pretend play have been documented in both chimpanzees and gorillas (Jensvold & Fouts, 1993; Ramsey & McGrew, 2005). The motivation for such behaviors is not obvious from the perspective of existing work on niche construction in the predictive processing framework, because these behaviors do not involve the identification and resolution of preexisting environmental uncertainties. Rather, we speculate that these behaviors could result from efforts to create uncertainty and surprise in environments in which they are lacking.

Interestingly, there is an apparent cross-species relationship between playfulness and the capacity for culture. Some of the most playful species, including dolphins, great apes, crows, monkeys, and, of course, humans, show highly diverse culturally patterned practices (e.g., Hunt & Gray, 2003; Kuczaj & Highfill, 2005; Whiten et al., 1999). While we recognize that this relationship is likely to be mediated by intelligence and general cognitive capacity among other things, we speculate that proneness to boredom and a proclivity to play may act as a creative stimulus for cultural innovation. Numerous researchers have already argued that human play facilitates creativity and innovation (e.g., Bateson & Martin, 2013; Russ, 2014). Whether this argument can be extended to other playful species remains to be seen. If these species modulate the environment so that surprises may be extracted from it, this could galvanize the emergence of new behaviors which, if they persisted over time and were transmitted between individuals, could be added to the cultural repertoires of their populations.

Play, Playfulness, and Mood

In addition to elegantly integrating emotion, cognition, and perception, recent predictive processing accounts have also emerged to include overall mood states into the framework (Clark et al., 2018; Kiverstein et al., 2020). Moods are often described as “generalized emotions,” emotions that are directed at the world as a whole rather than any one particular object (Solomon, 1993, 71). Moods are further distinguished from emotions by being longer in duration, providing a persistent “background” feeling tone to our transitory, short-lived emotional experiences (Ekman & Davidson, 1994). Like feelings, moods are also believed to structure our experiences by way of anticipation-fulfillment dynamics (Kiverstein et al., 2020; Ratcliffe, 2008).

From a predictive processing perspective, affective valence acts as a metacognitive signal within the predictive system, informing it of how well or poorly it is predicting in some specific local context. Moods by contrast are global background expectations about the slopes of error reduction the agent is likely to encounter. A positive mood, then, can be understood as the product of a series of experiences where the organism has reduced error faster than expected. This in turn leads to a general upward biasing of our expectations of positive valence going forward. In other words, agents that are in a good mood expect error slopes to incrementally improve (Kiverstein et al., 2020; cf. Clark et al., 2018; Eldar et al., 2016, 2021; Rutledge et al., 2014).

Playfulness has previously been recognized as a positive mood state that is frequently manifested in observable behavior during play (Bateson & Martin, 2013). While this mood state is believed to often accompany play, it is also believed to sometimes facilitate it. In the predictive processing theory of mood, repeated experiences of better-than-expected error slopes improves mood (Rutledge et al., 2014), making the agent more optimistic, and expect attractive opportunities to reduce error (Cools et al., 2011; Niv et al., 2006; Somerville et al., 2013; Wang et al., 2013). This is supported by laboratory findings that positive mood has been shown to induce risk-taking behavior (Arkes et al., 1988; Isen & Patrick, 1983) as well as in real-world settings (Bassi et al., 2013; Edmans et al., 2007), in which positive mood has been shown to bias the expectation of future positive outcomes (Wright & Bower, 1992).

Notice the effect that this positive biasing can have in an environment like ours where opportunities for error reduction tend to rise and fall together. The upward biasing of the agent’s expectations about the rate at which error is reduced makes it more likely for the system to expend energy to confirm predictions about error reduction slopes (Eldar et al., 2016). The optimistic agent is therefore more likely to find better than expected opportunities in their environment when they are available, which in turn perpetuates the positive mood. In this perspective, moods reflect a sort of emotional “momentum”—when the agent feels rewarded for doing better than expected, it increasingly expects such rewards to keep on coming (and conversely, when agents are doing worse than expected, it incrementally expects more bad times ahead, Eldar et al., 2016, 2021; Kiverstein et al., 2020; Rutledge et al., 2014).

Consider a child who has previously enjoyed a visit to a theme park. At the theme park, the child repeatedly experienced reducing error faster than expected where what is expected relates to the child’s preferred states, the satisfactions of its needs and desires, and the fulfillment of its goals (e.g., eating ice creams and candy floss that increase glucose blood levels faster than expected; the roller coaster rides that create and resolve error faster than the family car). That child is likely to become in a good and playful mood when being told that the family again this year is going to visit the park on the weekend. According to the model, this is because the child anticipates encountering a plethora of attractive error reduction slopes when reaching the theme park. In other words, the child is in a good mood then because it expects to encounter rewarding possibilities and the good mood will be sustained as long as this expectation is fulfilled. Mood is therefore a form of generalized summary of expectations that relates to how well or badly the agent has been faring in the world as a prediction error minimizing organism, which in turn shapes its anticipation of the trend of rewards going forward.

However, as the opportunities to reduce error begin to fall away, as will inevitably happen in an environment offering finite resources, the agent’s positive mood will likewise diminish. The theme park, for instance, offers a rich abundance of opportunities for the child to fulfill their desires until the park closes and a long drive home awaits them. Many studies suggest that a negative mood is associated with biasing of predictions for negative error slopes—anticipation of doing worse than expected in error reduction, biasing perception of negative outcomes (Badcock et al., 2017; Fabry, 2020; Kiverstein et al., 2020; Kube et al., 2020; Paulus et al., 2019; Ramstead et al., 2021). For instance, in depression, a state characterized by a persistent negative mood, there is a loss of confidence that any policy will succeed in reducing error (Badcock et al., 2017). This sometimes creates a perpetuating negative spiral, where the expectation of encountering worse than expected slopes for error reduction leads the agent to sample the environment for evidence, which in turn confirms and supports the negative belief. In that sense, playfulness as a mood can be thought of along the same lines as the famous words of Brian Sutton-Smith, who stated that the opposite of play is not work; it is depression (Sutton-Smith, 1997, p. 198)

Further Implications and Future Directions

The central role of positive valence in a predictive processing account of play may provide important new directions for future studies. Methodologically, it implies that zooming in on surprise dynamics over time may allow play researchers to get an important and empirically well-founded picture of the cognitive and physiological fluctuations that happen when children and adults engage in playful activities. At the same time, this may also provide play researchers with an alternative to unrefined between-group designs, given that surprise by definition is a reflection of the knowledge of the given participant. The framework’s emphasis and focus on predictions and prediction errors may lend itself to an increased focus on within-subject measures of agents’ real-time patterns of prediction on various time scales in different play settings. Recent technological advances may help here. Mobile eye tracking, for instance, is a particularly strong candidate for gathering behavioral proxies for predictions in ecologically valid playful situations (e.g., Andersen et al., 2019), and pupil dilation has been shown to signal uncertainty and surprise (e.g., Lavín et al., 2014).

Some of these methodological approaches are already widespread in the study of infant cognition, but grow increasingly absent in research paradigms as children acquire language and motor skills. Indeed, one of the most widely used approaches to study infant cognition has been to treat surprise or its absence as the main measure by observing whether children express expectation or surprise in various experimental contexts (e.g., Baillargeon et al., 1985; Scherer et al., 2004; Werker et al., 1997). Using such measures, efforts to systematically map what types of predictions children make in different forms of play could prove beneficial. Other behavioral measures can serve as proxies or indicators for surprise as well, and one could, for instance, use pitch levels in verbal utterances (e.g., Ververidis et al., 2004) or facial expression (e.g., Cohn et al., 1998) as behavioral proxies for surprise. The same goes for physiological measures such as heart rate variability (Andersen et al., 2020; Sukalla et al., 2016), which nowadays can be easily measured in real-life settings in noninvasive ways.

Future studies may benefit from tracking the relationship between various slopes of actual and expected surprise reduction over time and their effects on valence and motivation in play. For example, researchers might utilize an unfamiliar toy type, such as a drone controlled by hand gestures, designed to respond to the playing individual with various levels of unpredictability controlled by the experimenter. By tracking the gaze and hands of the playing individuals as well as the ongoing changes in distance from the hands to the toy, which is already possible with available technology, researchers will be able to get measures of how well participants predict the movements of such a toy over time and how predictions improve or worsen. Such measures may be then related to measures of interest, such as enjoyment or motivational measures, which could be obtained by showing the participant the first-person view video of their play and continuously rating it for how fun or engaging it was.

Simpler setups may also work. For instance, Doan and colleagues presented 4-year-olds with a puzzle that they were told was either easy or hard (or, in the baseline condition, where they received no information). When the 4-year-olds completed the puzzle that they were told was hard (i.e., presumably completed the puzzle faster than expected), they spent more time exploring and attempted more different interventions with a subsequent novel toy compared to when they were told that the puzzle was easy or at baseline when no difficulty information was provided. Thus, experimenters may take advantage of the possibility to manipulate the relationship between expected and actual surprise reduction over time. They could also investigate the effects of encountering several such instances, where agents do better than expected, which is hypothesized to positively affect their playful mood and overall risk taking.

For pretend play, researchers may consider taking advantage of the rise in popularity of online streamed tabletop roleplaying games, where older children and adults pretend to be characters in fictional settings. Through the use of automated voice recognition software, some of which is already implemented in larger online platforms (e.g., YouTube), researchers have access to vast datasets of dialog in the form of subtitles from pretend settings. Through the use of natural language processing (NLP) methods (e.g., Jurafsky & Martin, 2000), it is possible to characterize the moment-to-moment development of variables such as novelty and recurrence, syntactic complexity and narrative arc, while relating these measures to proxies for enjoyment like popularity, view count, or positive sentiment in language use in viewers of the stream. This, in other words, allows researchers to look at proxies for surprise/renewal and enjoyment and their intertwined relationships as they unfold over time whilst being completely in the sphere of imaginary forms of play.

Future studies may also investigate how the seeking, creating, and resolving of error slopes in play is mediated and modulated during playful interactions with other agents. We know from other studies of playful parent–child interaction, for example, that parents actively guide and manipulate expectations by signaling surprise to their children at appropriate moments. Mothers of toddlers have been shown to increase their mean fundamental frequency and use a wider pitch range in playful situations compared to nonplayful situations (Reissland & Snow, 1996). Similarly, another study involving mothers and infants interacting together with a surprise-inducing toy found that the mothers’ exclamations of surprise became more high-pitched when they noticed that their children did not react with surprise to the toy (Reissland et al., 2002). Along similar lines, Wu and Gweon (2021) introduced 3- to 4-year-old children to a novel toy with one salient casual function that the children first learned about. The children then saw an adult play with the toy. Intriguingly, children explored the toy more when the adult expressed surprise compared to when she expressed happiness, but only when the children knew that the adult already knew about the toy’s salient function. As Wu & Gweon argues, these results suggest that “children consider others’ knowledge and selectively interpret others’ surprise as vicarious prediction error to guide their own exploration” (p. 862). Thus, it may be that when agents have fun together, they do so by collaboratively reducing error for each other.