Chimpanzee play sequences are structured hierarchically as games

Sample

We scanned 116 h of video material from the Bossou video database (Matsuzawa & Humle, 2011), collected between 2009 and 2013. While footage from the Bossou outdoor lab has high video quality and filming consistency, the social composition of the group limits generalisability. The Bossou community at the time was small (around 13 individuals) (Matsuzawa & Humle, 2011). Due to the age distribution, there was only one infant, one juvenile, and one subadult individual in the group during data collection—making it difficult to differentiate between age effects and individual preferences (Fröhlich, Wittig & Pika, 2016). Eleven individuals were observed playing at least once; however, the distribution of observations was highly skewed, with the two juvenile/subadult players each participating in about 75% of all play bouts, while none of the adults participated in more than 20% of play bouts. Thus, most play elements and transitions were provided by two individuals, often playing with each other. In this study, we do not control for individual or age differences in play behaviour and sequences, due to the limited sample. These could make play transitions more predictable (individuals or specific age groups might have standardised ways of reacting that other group members know). Considerably more data would be necessary to control for individual- or dyad-level effects in transition patterns. We identified 143 bouts of social play across 35 videos—defined as play involving at least two individuals, with a new bout started if both individuals stopped playing for at least 5 s continuously (Liebal, Call & Tomasello, 2004). Bouts consisted of between three and 181 individual play elements (mean = 30.3), including between two and four players at any given time. For analyses, the bouts were split into individual-level bouts (every play element an individual performed during a bout), resulting in 306 individual-bouts.

Coding scheme

The coding scheme, with detailed definitions of all play elements and coding conventions, can be found in the associated repository. Potential play elements were identified from several sources—primarily, every behaviour indicated in Nishida et al. (2010) as potential play behaviour, the literature on ape gestural repertoires (Genty et al., 2009; Graham, Furuichi & Byrne, 2017; Hobaiter & Byrne, 2011; Hobaiter & Byrne, 2014), previous chimpanzee play literature (Fröhlich, Wittig & Pika, 2016), and descriptions of play elements in primates more widely (Petru et al., 2009). Often, these sources use different terms for similar play elements, and the definitions used here do not always overlap perfectly with those used previously. To our knowledge, the ethogram used here is the most detailed ethogram for chimpanzee play to date. Play elements can roughly be categorised as contact or non-contact, and as events (countable, one-off or repeated actions) or states (continuous behaviour with a clear start and end point). Social object play formed its own category, with multiple different ways of interacting with detached objects (mainly stones, nuts, and sticks) available. We coded play faces and vocalisations (where audible) when they were clearly visible—however, the distance and video quality did not enable us to code facial expressions reliably, and we did not include their placement in the analysis. In total, our ethogram contained 118 different play elements, of which 106 were observed at least once. We assumed that the elements we defined are meaningfully different from each other. This might not be the case: the difference between Retreat (walking away from partner), Flee (running away from partner), and Retreat Backwards (walking away from partner while looking at them) might be an artifact of the coding scheme.

Coding was done using BORIS v.7.9 video coding software (Friard & Gamba, 2016). We coded bouts one player at a time and marked the start of every change in play element and mark all active play elements at that time point. For example, if an individual goes bipedal, this is marked. If, while bipedal, the individual approaches the partner, we would mark bipedal/approach. If they would then raise their arm while performing those actions, we would mark bipedal/approach/arm raise, and so on. This leaves us with a string of play elements with a time stamp for initiation. If any player stopped playing (i.e., no play element was active), a Break was coded. The duration of play elements was available but was not considered in this study—we focus entirely on the sequential order.

Video coding of entire play bouts is slow, due to fast changes of behaviour and movements, and researchers usually focus only on play initiation and re-initiations (Heesen et al., 2021a; Hobaiter & Byrne, 2011). Due to the challenges of this detailed coding approach, no inter-rater reliability was performed, and results must be viewed with this limitation. Predictability should be higher in studies using simpler coding schemes, so if we can show high predictability using the current ethogram, we have taken the conservative approach. The dataset currently contains 5,711 play elements. Where possible, we present results including uncertainties, and used permutation and bootstrapping approaches to discriminate between spurious and reliable transition patterns.

Pre-processing

All pre-processing and analyses were conducted in R statistical computing software (R Development Core Team, R Core Team, 2020). The video coding data needed pre-processing to deal with three main problems inherent to the coding process: rare elements; some artificially common elements; and establishing the sequential order of co-occurring elements.

To robustly establish probabilities of transitions between elements, rare elements are a problem (Silge & Robinson, 2017). For example, if an element only occurs three times, and each time transitions into a different element, we do not know if the high transition probability would disappear with increasing sample size. We set the threshold at 20 occurrences per play element. However, removing these cases completely (as is often done in linguistic studies; Silge & Robinson, 2017) would be wasteful given the sample size of this study. For most play elements, we defined a priori with which other play element they would be combined if too few occurrences were observed (see Supplemental Information). Replacement elements were chosen based on similarity of movement. If the combination after this lumping process failed to reach the threshold, we nevertheless retained it. Thus, our rarest element had nine occurrences (see associated repository for occurrence probabilities of all play elements before and after pre-processing). After this step, 68 play elements remained.

Some elements occurred at much higher frequencies than others. The seven most common elements (Bipedal, Hold, Follow-Other, Approach, Retreat-backwards, Retreat, Flee) were all coded continuously and therefore were noted every time a change occurred while they were active. Imagine a musical piece on the piano: sometimes one note is held while others are played. In play, a chimpanzee could go bipedal, but then perform other actions while the Bipedal was marked at every change in event. These elements potentially skew transition probabilities and mask transitions between other elements. Ideally, we want a sequence that reflects when individuals made the choice to use a specific element. We addressed this by detecting cases where one of those seven elements occurred multiple times in a row, and only retained the first case. If players stopped the continuous action (e.g., stopped fleeing, then started again), the element was counted again.

In play, it is possible to go Bipedal, Arm Swing with one arm and Hit the partner with the other arm. This is problematic in terms of the transitions - does Bipedal lead to Arm Swing; or Arm Swing to Bipedal? This problem also occurs because some elements (e.g., Flee or Bipedal) are continuous states, while others (e.g., Kick) have a clearly defined beginning and end. We used permutations—randomly assigning order within co-occurring elements and repeating all analyses 1,000 times with different orders—as there was no a priori reason to assign primacy to one co-occurring element over another. Thus, all described transition probabilities are averages over multiple permutations, which is why transition counts are not integers. Two alternative approaches (random sampling of only one of the co-occurring elements, bag-of-words) can be performed using the attached R scripts and generally showed similar results.

Transition probabilities

The transition probability between antecedent and consequent were defined by the number of times the consequent followed the antecedent, divided by the number of times any element followed the antecedent (conditional probability). The antecedent could be a single element (used to establish first-order n-grams, networks, and transition similarities), but also n-grams of different order (e.g., first order: Hit; second order: Hit/Slap; third order: Hit/Slap/Tickle etc.). The latter approach was taken to determine whether increased information about antecedents increases prediction accuracy. Current sample size prevents us from analysing long sequences, as the number of possible transitions increases exponentially with each new level. We limited the analyses to a maximum of three antecedent elements. We restricted ourselves to one-element consequents and did not consider non-adjacent contingencies (Sonnweber, Ravignani & Fitch, 2015; Watson et al., 2020).

The large number of possible combinations combined with a small dataset and the small number of individuals leads necessarily to overfitting: some combinations will only occur a few times and adding new information could influence our understanding of their function. We did two things to counter this: rare elements were combined, as described above. Where possible, we report some measure of robustness to give the reader an understanding of how reliable results were. Robustness was established using bootstrapping procedures—randomly selecting 1,000 subsets of the data and establishing transition probabilities within those subsets.

Randomisation procedures

To test which elements followed which antecedent, we created a null model of ‘expected’ transitions using permutations of observed patterns. We chose this resampling approach over collocation analysis (Bosshard et al., 2022) to account for the regular co-occurrence of play elements that is usually not seen in single-modularity communication; however, using collocation analysis instead did not alter the results. We repeatedly randomized the order of elements across bouts: while the number of elements per bout, the probability of elements to occur across bouts, and the position of Breaks and missing data in each bout were kept the same, we randomly assigned element positions. Thus, transitions are considered significant if they were observed more often than would be expected if play elements were just strung together randomly given their base probabilities. We ran 1,000 randomisations to create the expected distribution for each transition and compare whether the observed transition probability fell within this distribution or not. To compare the observed and expected values, we provide a p-value (how many of the 1,000 randomisations show higher transition probabilities than observed; Mielke et al., 2021). We report transitions that occurred at least five times and that were significant at 0.01 level (i.e., the observed value was higher than for 990 permutations). These calculations also constituted the basis for the network clusters described below.

Prediction accuracy

To understand the predictability of transitions rules, we applied the probabilities derived from a subset of the data to ‘unknown’ test data and explored how well the former predicted the latter (Chollet & Allaire, 2018). We tested the predictability of elements within bouts by calculating transition probabilities for 95% of all other bouts, then predicting each element in the remaining 5% of bouts based on their antecedents (k-fold validation). This was repeated 1,000 times per bout. We tested the expected correct classification if the consequent element was only determined by base occurrence probabilities (null model). The difference between this value and the observed prediction accuracy of the models tells us how much knowledge of the antecedent increases our predictions. Aside from using one element as antecedent (describing a simple Markov process), we repeated the process with two or three elements as antecedents (n-gram prediction; Eisenstein, 2019). For higher-order antecedents, the probabilities of the lower-order antecedents were combined (interpolation)—therefore, for Approach/Stare At/Hit as third-order antecedent, the probability is the product of the probabilities of the triad, Stare At/Hit, and Hit. This was done because many higher-order antecedents only occurred infrequently, and no information would otherwise be available as to which consequent was appropriate. For transitions that were never observed, Laplace smoothing was applied, assigning them one occurrence, and adapting all transitions accordingly (Eisenstein, 2019). If the prediction accuracy under those conditions was higher than for one element, this indicated hierarchical processes - for example, if Hit correctly predicts to Hold 10% of the time, but Stare At/H it leads to Hold in 80% of the time, then the sequence order added information. We present the mean correct classification rate across all bouts and elements. In addition to predictions based on the transition probabilities, we implemented a naïve Bayes classifier using the ‘e1071’ package in R (Meyer et al., 2021). Naïve Bayes classifiers use vectors of feature values (in our case, the previous play element, two previous play elements, etc.) to predict the correct consequent using Bayes theorem (Eisenstein, 2019). Using an established classifier offers the advantage that classification is optimised and faster than the above-described prediction based on raw transition probabilities. However, naïve Bayes classifiers make a strong independence assumption, effectively assuming that the antecedents are independent from each other given the consequent class (Eisenstein, 2019). Therefore, while increased performance of the classifier with increasing number of antecedents would indicate that information about previous play actions increases predictability of what happens next, performance cannot be interpreted as based on sequential information.

Similarity

We determined whether there were play elements that resembled each other in which elements followed them and tested whether we could find clusters of similar elements. This is similar to the identification of synonyms in language (Levshina, 2015), and we did it both to test whether our assignment of distinct elements during coding was meaningful and to see whether there were clusters of interchangeable elements. Each element was represented by a vector of transition probabilities with all elements. We applied Uniform Manifold Approximation and Projection (UMAP; McInnes, Healy & Melville, 2018) to achieve two-dimensional representation for each vector using the ‘umap’ package (Konopka, 2022). We established similarity between play elements by calculating the Euclidean distances between UMAP projections. To identify the optimal number of clusters for the hierarchical clustering, we used K-Means clustering as implemented in the ‘cluster’ R package (Maechler et al., 2022) to determine (a) the optimal number of clusters, and (b) the quality of the cluster solution. We present the silhouette value (Rousseeuw, 1987) to detect the best cluster solution, indicating an acceptable distance between clusters and coherence within clusters—any solution above 0.3 can be considered to show that there is more similarity within than between clusters. As cluster solutions differ based on the outcome of the UMAP dimension reduction, we repeated the dimension reduction and cluster detection 50 times with varying numbers of epochs for the UMAP (on average 7000 epochs) and continue using the most likely cluster solution. We plot the dendrogram for the optimal cluster solution and saved cluster memberships for later comparison with network clusters.

Networks

Networks can be useful tools to visualise the connections between elements in communication networks and to identify clusters of elements that have above-expected connections with each other (Allen et al., 2019; Aychet, Blois-Heulin & Lemasson, 2021; Barceló-Coblijn et al., 2017; Mielke et al., 2021; Weiss et al., 2014). Here, we created a network using all play elements as nodes and the transition probabilities between them as edges (Newman, 2010). Only transitions that were significantly more likely than expected and occurred at least five times were considered, to make the network intelligible despite the large number of elements and ensure biological relevance. Edges were weighted, representing the transition probabilities between elements; and directed, meaning that each dyad of elements was represented with two values (A to B, B to A). We used the ‘igraph’ and ‘ggraph’ R packages (Csardi & Nepusz, 2006; Pedersen, 2021) to create and visualise networks. To test whether distinct ‘clusters’ of play elements existed in the network (indicating groups of play elements that have strong connections with each other but weak connections to the outside), we used the ‘cluster_optimal’ community detection algorithm in igraph, which maximises modularity of clusters (Csardi & Nepusz, 2006). Clusters were considered to represent stronger connections within than between clusters if the modularity value of the cluster solutions was larger than 0.3. Cluster solutions were compared to those produced by the similarity measure above.

(a) Non-random transitions

There were 1,622 transitions that were observed at least one time. The histogram (Fig. 1) shows that most elements are followed by several different consequents with low probabilities. In only four cases did a consequent constituted more than 30% of all possible transitions of an antecedent, with two of those (Drum Tree and Kick Dirt) being loops—the element was repeated sequentially. At the same time, each element was observed to be followed by between seven and 53 elements. Thus, there was no tight coupling between any two elements. This might indicate random assignment—any elements could be followed by any other. However, it might also mean situation-specific responses that were tailored to the players’ own previous action and the partners’ reaction, or predictability at a higher order (e.g., based on multiple antecedent).

We also visualized how robust transitions were (Fig. 2). Using bootstraps, we created an interval around the observed transition probabilities. We plotted the range of values for each transition for the 1,000 bootstraps (calculated as the highest transition probability minus the lowest transition probability of A to B in the set) against the number of times the antecedent was observed. For some rare elements, transition probabilities remained volatile. Transition probabilities of rare elements therefore must be interpreted with caution, and elements will be filtered to exclude rare transitions—in all descriptions of ‘significant’ transitions and in the networks, only transitions that occurred at least 5 times were considered and reported.

In total, 146/1622 transitions (9%) were significantly more likely than expected. More detailed depictions of these patterns can be seen in the network below and in the associated repository. When analysing the non-random transitions in detail, we found that many elements significantly followed themselves (21 out of 147 significant transitions). Several of the elements used here—for example, rocking or drumming on an object—are repeated actions and each occurrence was marked as independent event. In contrast to all observed transitions described above, many elements (17/68 elements) had no significant consequent, 14/68 had only one significant consequent, with the maximum number of significant transitions in one antecedent being 10 consequents (for Holding the partner and Bipedal).

(b) Next-element predictions

When applying the transition probabilities as predictions, increased information about antecedents increased predictability (Table 1). The basic probability of correctly predicting an element based on its occurrence probability (zero-order) was 0.03. By applying the probability of one antecedent (unigram; e.g., Hit) we increased the probability to 0.06—almost a doubling of correct classification. When adding two antecedents (bigram; e.g., Bipedal/Hit), there was another rise to 0.11—again, almost a doubling of correct classifications, and almost four times higher than having no information about antecedents. At the third order, we do not achieve further improvement. For the naïve Bayes classifier, using a more optimised approach that however assumes independence of antecedent elements, we achieve correct classification results of 0.07 as baseline, 0.13 for the first order, 0.16 for the second order, and 0.14 for the third order. Thus, additional information about preceding elements improved prediction accuracy. However, there was still a lot of unexplained variation.

(c) Similarity between elements

In Fig. 3, we can see the dendrogram representation of hierarchical clusters of distances between transition probability vectors of all play elements. Elements connected through shorter branches and assigned the same cluster membership (same colour of branches) are considered more similar than those further away and with different colours. The best cluster solution, with silhouette value of 0.68 (indicating a well-distinguished cluster solution) contained 12 clusters. The cluster allocation can be seen in Table 2, and we will discuss their potential classification together with the network. What we can see here is that there were many elements that were similar in consequents. For example, Kicking the partner and Jumping on them were close, indicating that they could have been defined as a single play element. Similarly, Retreating Backwards and Retreating were closely connected. A lot of similarity between elements can be explained by their frequent co-occurrence—for example, Retreat and Bipedal showed high similarity because chimpanzees often retreat from the play partner while bipedal.

(d) Network structure

In contrast to the similarity clusters, which assess whether two elements are used at similar points in a sequence, the transition network (Fig. 4) describes which consequent follows which antecedent. The network only depicts transitions that occurred at higher-than-expected rates and occurred at least five times in the dataset. Colours indicate community membership. As the high modularity of the network community detection algorithm (modularity = 0. 65) indicates, there were seven clearly distinguished communities in the network. If community assignment was random, we would expect around 32% of transitions between the elements within communities, but we observed 48% of transition within communities—a 1.5-fold increase. Connections between communities were often due to elements that can be used in different situations. For example, Shake Off is used when playing wrestling with a partner to get away, but equally when the player is hanging off a branch or retreating—hence, the element is connected to three communities. Individuals stomp when initiating play in combination with Bop and Bow, but also when they were bending a small tree and holding onto it.

For the interpretation of communities, in combination with the similarity clusters, see Table 2. There was considerable overlap between the two approaches, with small variation arising mainly because several elements did not have any significant transitions above threshold level, and the combination of object-related and movement elements resulted in overlap between the chase and object clusters. The different cluster combinations (‘games’) can be categorised broadly by whether they involved climbing by either partner, had physical contact, involved chasing, involved objects, or were play invitations. For the latter, one clear cluster emerged, consisting of Bop, Bow, Stomp, and Slap Ground, which individuals often combined and repeated in quick succession to indicate that they were willing to play. Some other, rarer elements (Present Body Part, Rock, Kick Dirt, Stare At) can fulfil a similar function. Play elements routinely used when one or both individuals were in a tree transitioned into each other at high rates, depending on the role of the focal individual. When the player was on the ground and the partner in the tree, individuals would often bend the tree (the most central element of this cluster), and then pull or shake it, sometimes while jumping. Players regularly hide swing (swinging around a tree at speed) before climbing up. While players were in the tree, they climb up and then hang while swinging, kicking the partner, shaking them off, and ultimately falling.

Most contact play formed one large community in the network, with elements transitioning into each other at high rates. Based on the similarity of transition probabilities, we could differentiate two groupings: contact play that involves players to stay in one spot (Bite, Wrestle, etc.), centred on holding the partner in place; and those that involve one player trying to get away from their partner while still in contact (Push, Trip etc.).

The different object-related play elements were connected, including detached objects and trees. Chimpanzee players held onto objects once they had grabbed them and then manipulated them in different ways. Object contact was the defining element of this type of play. A common way for the Bossou chimpanzees to initiate play with object contact was to roll objects towards the partner or press the ground. Individuals will often drum trees. Players wave objects about while swaggering towards the partner and flailing or waving their arms. Many of the social object elements were connected to retreating movements, with the player retreating while holding an object, which explains the community overlap of object interactions and avoidance movements.

The remaining cluster combinations were related to chasing play on the ground. Again, we can identify different roles of the player: One community were those elements strongly connected to movements used to avoid the partner retreating or retreating backwards from them and hiding behind trees or feinting directional changes, often lifting their arm protectively. They will circle the partner while parrying hits. The last cluster combination involved the opposite, with the individual approaching the partner (sometimes following a pirouette as play initiation, often combined with bipedal movements and arm swings), chasing, and trying to make physical contact while the partner flees (Reach, Hit Attempt).