``Backseat Gaming" A Study of Co-Regulated Learning within a Collegiate Male Esports Community

2.1 Learning in Esports

In the context of learning to play games, the end goal is often expertise. Despite expertise being recognized as a key component of gameplay [21], studies of expertise in games are often more interested in the skills possessed by expert players, rather than how those skills are learned. Nowhere is this more apparent than esports research, which often discusses and formalizes the skills experts use to succeed in gameplay [20, 24, 37, 48, 66] but has been, overall, less interested in studying how novice players gain these skills over time. As Kleinman et al. pointed out [39], this gap in the literature makes it difficult to design computational support for the learning process, as we do not possess a strong understanding of what that process is.

In an attempt to address this gap, Hesketh et al. [34] leveraged grounded theory to examine learning in team esports and identified several preliminary themes around the topic, including several related to how players had to evaluate and adapt their skills to new scenarios. In another example, Pluss et al. [60] performed a larger scale study determining the connections between amount and quality of practice over time and ultimate tournament performance. In doing so, they highlighted how just playing a game did not correlate directly with learning and performance gains and that there were other factors to consider. Building on these findings, Kleinman et al. [39] conducted an interview study specifically investigating the activities that esports players leveraged to try and gain skill in their chosen game and the challenges they faced in that process. They took the topic a step further, however, using their results to inform a set of suggestions for future computational support tools to better meet players’ needs.

While this work has been critical at filling the gap in our understanding of esports learning, it is also noteworthy that both Kleinman et al.’s [39] and Hesketh et al.’s [34] participants emphasized that their learning processes often involved obtaining information from other sources be it known others or community resources. Neither work, however, places a strong emphasis on exploring this topic, instead suggesting it should be examined further in the future. An older study by Kow and Young [44], however, does look into this phenomenon in far more detail. Their interview study with StarCraft community members revealed the critical role that both personal and community resources play in, not only motivating players, but helping them advance their skills and keep up with changing expectations within the gameplay community. Through this work, they illustrated the central role other people play in an individual player’s learning process, as a means by which knowledge can be leveraged towards one’s own learning goals.

2.2 Social and Co-Regulated Learning in Esports

This leveraging of the knowledge of a more experienced “other” individual (be it a specific person or a community resource) suggests the presence of Co-Regulated Learning (CoRL) [31] practices within esports communities. CoRL is a theory of learning that branched off of Self-Regulated Learning (SRL), suggesting that SRL skills are acquired from other, more experienced individuals through a process of appropriation [30, 31]. Similar processes have been observed in esports teams where new players are mentored by their experienced counterparts [71]. However, as Kow and Young pointed out [44], when it came to leveraging knowledge and learning over time, expert players were far more likely to be able to interact with other people in the process, leaving novice players reliant on public media technologies, community resources, and, likely, computational support tools. Kleinman et al.’s participants [39], however, indicated that existing computational and media technologies cannot replace a real coach or teammate in the learning process because they do not think or act like one. This prompted them to suggest that the research community needed a stronger understanding of CoRL in esports to design better computational support tools for the domain.

CoRL, and even social learning as a process that occurs over time, is, however, understudied in esports. In fact, much of the work on social esports play is more focused on elements of teamwork and how they inform performance and victory over how team-members help one another learn [55, 57, 62, 78, 79, 88]. Instead, observations of social learning are often included as a side product of a research study otherwise focused on solo learning or a different topic altogether [39, 67]. Richard et al. [69], in their study of collegiate esports players, do extensively examine social learning activities, including group strategy planning and reflection. However, this study looks specifically at how these activities are engaged during a single competition, and does not examine how input is exchanged and learning is guided by others during the long-term process of gaining expertise in the game. While it provides valuable information about team-based metacognitive processes, it does not tell us much about how these are leveraged over time to help each other learn.

Rusk et al. [71], in contrast, examined social learning over time through mentor-apprentice relationships in esports through a qualitative study of a Counter Strike: Global Offensive team. Their findings indicated how new players were integrated into the team and provided with guidance and support by teammates until they demonstrated enough competence to be considered an individual agent on the team. Unlike the previous example, Rusk et al.’s work does explore how learning occurs long term within an esports context, and in doing so captures how learning and relationships change throughout the process. One of their most interesting findings is the apprentice player’s shift from obedience and excuses to resistance to input and justification of choices once they have achieved enough skill to act independently. Rusk et al.’s work is perhaps one of the most in depth explorations of how social learning occurs over time in esports contexts, with the above paper being one of several examinations of the topic of mentor-apprentice relationships [70]. However, they studied a team with a clearly identifiable apprentice and mentor and, despite looking at how the mentor helped the apprentice learn, did not leverage CoRL as a theory in their work. Here, we build upon this previous work by examining social learning through the lens of CoRL and within a community where the roles of apprentice and mentor are not so clearly defined. In doing so we hope to expand our understanding of social learning in esports and provide actionable insights for the advancement of computational support.

2.3 Computationally Supporting Learning in Esports

Some of the most prominent computational support tools for esports are visualization systems that have been developed to provide players with gameplay data in a human-readable manner to better facilitate learning processes such as reflection [2, 84]. Wallner et al. [85] even conducted a user study examining how players use visualization systems to learn and what features were most useful to them. An example of one such system is Kuan et al.’s [45] spatio-temporal system for Starcraft 2 synchronizes data presented across multiple maps and timelines. The user can highlight interesting spans of time in the timeline, which controls what is shown on the map, allowing them to reflect on their gameplay, identify errors, and develop ways to adapt their strategy in the future, all elements of learning discussed in prior work [38, 69].

There have also been similar systems designed for spectator use. For example, Charleer et al. [10] designed a spectator dashboard for League of Legends that presented how much gold each player had at all times, as this was indicated by the player community as an important indicator of victory. Another example is the WEAVR application [41], an esports spectatorship visualization system that can be used by audience members during DotA 2 tournaments to follow the action, identify moments of interest, and anticipate what is to come. These systems help spectators learn from others by better understanding others’ decision-making processes and also helps novice or less experienced players become motivated to improve by watching more experienced players, an element of learning observed within existing esports communities [44].

While these visualization systems emphasize supporting reflective processes in both players and spectators, algorithmic systems that can predict what comes next and give players intelligent recommendations are seeing increased interest as a way to support planning and execution-related learning processes, either before or during play [11, 12, 22]. For example, the work of Semenov et al. [74] tested the accuracy of different quantitative methods (Naive Bayes classifiers, Logistic Regression, and Gradient Boosted Decision Trees) on predicting DotA 2 match outcomes based on what heroes were picked during the draft. Predictive systems like this are the foundation of tools that can provide players with in-the-moment information that can help them make informed decisions and learn during play by better connecting their actions to the outcomes they experience [38]. For example, Christiansen et al.’s [15] statistical approach to predicting chances of victory based on actions taken was later turned into a plugin for DotA 2 that displayed to the player what their chances of victory would be if they chose to pursue any of several next possible actions. Prior work has demonstrated that giving recommendations or discussing possible strategies in this manner is a common element of social learning in team esports contexts [38, 44, 69] and by providing this information to players, these tools are taking the first steps towards emulating the behavior of a coach or more experienced teammate.

Despite the success of prior work and the prevalence of commercially available tools [5, 54, 61, 75], there is still much to be understood about social learning in esports if these tools are going to meet the challenge of teaching like real people and make esports more accessible to more players. Here, we take one such approach to this challenge by leveraging CoRL as a theoretical lens to examine social learning via the exchange of input among an esports community. In doing so, we enrich existing literature by connecting esports learning phenomena back to an existing framework of learning and identify pathways for future work to explore and opportunities to advance the state of the art of computational tools.

3.1 Interview Design

We developed a set of questions for semi-structured interviews targeting input exchange practices among esports teams based on the theory of CoRL. The interview structure consists of three parts, with parts two and three informed by and targeting the two input related constructs of CoRL, as seen in Table 1. Part one of the interview asked for general information about opinions and habits surrounding social play. These questions, seen in Table 1, were not directly linked to any theoretical aspect of CoRL and were included to gauge how participants felt about social learning, what their experiences were with the phenomenon, how often they engaged in social learning and with whom.

We chose semi-structured interviews as this approach would allow us to focus on a single community at a time, ensuring that communities composing of minority groups explored via the same method in future work would be properly represented, rather than drowned out among a large scale examination. By using theory to derive the interview design we ensure that we are focusing on and targeting the phenomenon, input exchange, that we are interested in, which exists within the much more complex context of esports play. The semi-structured interview approach further allowed us to capture specific phenomena that may not be directly observable (i.e. emotional states) and ask follow up questions in response to anything interesting or unexpected said by the participants.

3.2 Interview Structure

Interviews were conducted in a one-on-one, in-person, semi-structured manner between October and December 2022. Audio was recorded and transcribed afterward. After receiving informed consent, the interviewer collected demographic information and then asked the questions seen in Table 1. During the interview, the interviewer also collected demographic information related to frequency of team play, frequency of input giving, and frequency of input seeking. All participants were asked all questions. The interviewer would ask follow-up questions or for elaboration if necessary. Participants were allowed to skip any question they did not wish to answer. Interviews ranged from 14 minutes to 41 minutes, with an average duration of 26 minutes. University IRB approved the protocol.

3.3 The Community

For this study, we focused on a competitive esports club at Abilene Christian University, located in the South Western United States. The club had 14 members at the time of the study, each of whom played at least 1 of 4 different games (see Figure 1) with some members playing multiple games. We chose this community because it is representative of a majority group within the overall esports community in the United States. As collegiate players between the ages of 18 and 22 the members of this community sit within the largest age group for esports fans in the US ⁴. While all male, the community reflects the generally male dominated esports culture where approximately 72% of fans identify as male ⁵ and fewer than 10% of professional players are female ⁶. The gender breakdown of the community is also consistent with samples seen in previous work on esports which also tend to be primarily or exclusively male [19, 51, 62, 82]. We discuss the implications of focusing on this community and the need to explore other demographic groups in limitations and future work. Race and exact age information were not collected to protect participant identities.

After each interview, the interviewer made note of the approximate themes that were discussed during the interview. Saturation [29] in these initial themes was observed around participant 10. After the remaining 4 interviews no new themes emerged during this initial review phase. The two researchers who conducted the data analysis, who were separate from the one who conducted the interviews, confirmed saturation in the data-set collected from the 14 community members.

3.4 Data Analysis

The data were analyzed via iterative thematic analysis by two researchers [27, 73]. In an initial pass, both researchers, separately, reviewed a representative sub-set containing 30% of the data [9] and developed initial themes. The unit of analysis was a single utterance by a participant. The researchers then reconvened to discuss their independent theme lists, collapse overlaps, and generate a combined list of themes containing 20 themes. The two researchers then performed an Inter-Rater Reliability (IRR) [16] check with the 20 themes. The resulting kappa value was.79, indicating strong agreement [47], however, the application results revealed three themes that were never applied by either researcher and three that the researchers never agreed upon, and each only applied once to the data set if at all. Another round of iteration on the code-book at this point saw these six themes collapsed into one another or other themes accordingly. This resulted in a list of 18 themes organized into three categories that corresponded to the three constructs that informed the interviews. One researcher then applied the 18 themes to the entire data set. The researcher then performed a final iteration on the code-book based on that application, collapsing and re-organizing themes based on number of appearances within the data set and overlapping meanings. At this point, one theme unrelated to the giving and receiving of input was also removed from the set. These were then re-organized into four categories based on the higher-level concepts the themes were related to. The final list consisted of 10 themes organized into four categories (See Results).

4.1 Demographics

The demographic data for the club can be seen in Figure 1. All participants reported having friendly relationships with their fellow club members. Most of the club members reported playing either Rocket League or Rainbow Six Siege as their main game. Both games, along with Valorant, also played by one member of the club, are included in the top-10 esport games for 2022-2023 ⁷. Years and hours of gameplay are in general terms, not exclusive to playing with teammates or with any specific game. Figure 2 shows reported frequencies of input giving and seeking and playing with others. These were collected via likert scale questions asking players “on a scale of 1 to 5 how often do you...".

Figure 1:

View Figure

Figure 2:

View Figure

4.2 Interview Results

An overview of the 10 themes and how often they appeared in the data set in terms of the number of total observations and the number of players who brought it up can be seen in Table 2.

5.1 The Hierarchical Nature of Input

The most prominent takeaway from the results is the emphasis on hierarchical relationships within the club and how one’s position in the hierarchy dictated their role in the co-regulation process. Notably, input was not always exchanged explicitly between club members, but when it was, those club members perceived to be less skilled were lower on the hierarchy and typically the recipients of input from those who were perceived as more skilled. The theory of co-regulated learning dictates how one who is more skilled may play an active role in a learner’s learning process by handling the metacognitive elements of the task (i.e. helping them reflect on what they did or determine their next move) and guiding the learner through those processes with appropriate input [31]. This system of having those with more experience guide those with less mirrors this phenomenon. Further, the existence of this hierarchical relationship is consistent with mentor-apprentice interactions discussed in previous work [70, 71]. However, while previous work examined a context in which the apprentice was easily distinguishable as a new player due to their time on the team, it appeared at first glance that players in our community all stood on the same level. Only through the interviews did we observe that the players still recognized a ladder within their community.

More interesting, however, is how serious the club members were about maintaining this hierarchy. Input was expected to come solely from those higher up and it was expected to be taken by those lower down without resistance, with players going so far as to suggest that failure to adhere to this arrangement could lead to conflict. In the context of using this work to inform the design of computational assistance, this echoes an earlier finding that players tended to not engage with tools they felt did not know the game as well as they themselves did [39]. This connection emphasizes the importance of building assistive tools based on an understanding of how players help each other (or allow each other to help).

The majority of the club members seemed to be aware of these rules, as most participants suggested that they would never even try to give input to those perceived as better than them. While CoRL suggests that this management from the more experienced is dialed back as the learner gains skill [30], our participants did not mention any such phenomenon, suggesting that the more skilled members of the club would always give input to those lower on the hierarchy, regardless of how much skill the lower player gained, and that the lower player was expected to accept and act on this input, not resist it. It may be that the strict adherence to a hierarchy prevented the players from perceiving those below them as ever having gained enough skill to not require their input. It may otherwise be that the more skilled players also believed themselves to be continuing to gain skill, and thus, while the players below them had gotten better, they too had gotten better, and their positions remained the same in relation to one another. This is in contrast to what is discussed in Rusk et al.’s work [70, 71], who found that support and orientation from other players diminished as apprentice players gained skill. This may be the result of cultural differences, as Rusk et al. examined Finnish students in their work, while ours were recruited from an American university, and prior work has demonstrated that gaming culture varies across the globe [18] as do interpersonal relationships and interactions [7, 72]. These differences highlight the need for further study of this topic with other demographic groups to expand our knowledge on the topic.

This strict adherence to a hierarchy is of additional interest given that participants suggested that there are contexts in which one’s status within the club may change. For example, a player who has a lower overall skill level may know how to play a specific character better than another player, which gives them a higher skill level in that specific context and permits them to give input, but only in that context. This flexibility comes at an interesting juxtaposition with the strict adherence to a hierarchy as it raises questions regarding how the club members recognized such contexts and shifts in the hierarchical relationship and what happened when there was a disagreement in relation to that shift. Furthermore, while players indicated their own strict adherence to this hierarchy, many had stories about someone else of a lower rank or status who had provided unsolicited advice. Often this other player would, in their own interview, suggest that they would never do something of the sort, and perhaps even accuse the one who had accused them (they were not aware of what was discussed in each other’s interviews). This suggests that the club members are aware of the hierarchy and their need to adhere to it, but may not always be aware of whether or not they are actually doing so.

5.2 The Relationship between Input and Failure

Another prominent trend in the results is the notable relationship between input and mistakes or failures. For the members of the club, recognizing a failure or the feeling that a mistake was made would prompt one to seek input while noticing repeated failure would prompt one to give input. It further appeared that being able to recognize one’s mistakes and seek out input was seen as a positive character trait or a sign of skill or growth by the club members. Previous work on CoRL, especially in computing education, saw a high level of social help-seeking behaviors, especially when hitting a wall [63], but CoRL occurring in response to recognized failures is not prominently documented or discussed. Rusk et al. [71], however, did discuss how expectations and responsibility were a major part of an apprentice player’s learning trajectory. They specifically discuss how expectations for the apprentice’s performance and the amount of responsibility they held grew as they gained skill. This may be the reason for the emphasis on failure we see here. As expectations grow, chances to fail do too, and as responsibility grows, the need to take responsibility for failures, by seeking input in order to overcome them, likely grows as well.

We further propose that this emphasis on identifying and overcoming failures is a side effect of input exchange occurring primarily pre- and post-play within the club. Post-gameplay, in particular, is referred to as self-reflection according to Zimmerman’s cyclical phase model and is characterized partially by the identification of mistakes and adaptation of strategies for future play [89]. While we discuss the timing of input further in the next section, here we suggest that post-gameplay input tends to gravitate towards this topic as a result of inherent SRL skills that the players may possess, possibly as a side-effect of playing the game, as suggested in previous work [37]. Further, previous work has discussed how difficult it can be for newer players to recognize their mistakes and identify how to prevent them in future play [38, 39] and it may be that the emphasis on examining failures helps the club onboard new members and move people up the ranks more easily.

5.3 The Three Phases of Gameplay and Learning

Consistent with previous work [37, 38], our findings demonstrate that learning processes occurred differently across the three phases of gameplay (pre-game, in-game, and post-game) during team play among the club members. Specifically, while our participants dictated that pre-game was dedicated to discussing strategies and setting goals and post-game was dedicated to reflection and evaluation, they did not discuss any input exchange in-game, during which time they focus on maintaining performance. Richard et al. [69], in their study of collegiate esports players, similarly found that group discussion pre and post (or between) games was similarly focused on planning and reflection and adaptation respectively and previous work [38] discussed how these three gameplay phases mirror the three phases of learning defined by the Cyclical Phase Model of SRL [89]. However, the previously leveraged Cyclical Phase Model of SRL does not account for social interaction, specifically the seeking or receiving of input within the learning process. Socialization was critical to learning within the esports club we focused on in this work and if this generalizes to other communities then it may be that the Cyclical Phase Model alone is not enough to describe learning in esports contexts.

In contrast, Hadwin’s CoRL model [31] proposes four phases: understanding, goal setting, working toward the goal, and adaptation. From our results, it appears that both phases one and two take place pre-game, while phase three occurs somewhat in-game and phase four is post-game. In phases one and two, the group negotiates their shared understanding of the task and their strategies for accomplishing it. In phase three, the group works collaboratively on their goal, collectively utilizing multiple cognitive, metacognitive, and motivational strategies. In phase four, the group makes small changes to large-scale pivots in strategies and goals based on feedback from the task and one another. This model may better apply to social learning in esports, as it better encompasses the critical role of social interaction and input exchange via communication between team-members, though it similarly downplays individual learning processes, which may still be important.

Further, given both of these models’ discussions of metacognitive processes occurring during the task, the lack of in-game learning within the club is interesting. While previous work also suggested that esports players are too focused during gameplay to think about learning [38, 39], it was proposed that they may not be aware of it. Here, the players more explicitly described that there is little to no input being sought or given in-game, as they wish to focus all in-game communication on the gameplay itself and keep negative emotions as low as possible to avoid them impacting play. At the very least, from this we can infer that there may be no social learning in-game within the club and that any subconscious learning that may occur happens at the individual level.