Finding winning team compositions by jointly modeling player proficiency and team congruency in multiplayer online games

In the messy, fleeting world of temporary teams, success often hinges on a single, difficult choice. When a group of strangers comes together for a short-term project, a weekend tournament, or a pickup game, they face a fundamental tension. Should each person take on the specific task they are personally best at, or should they fill the roles the group needs most to function as a whole? These two goals do not always align. A team might be filled with individuals who are experts at their own jobs, yet fail because their skills overlap too much or leave critical gaps. Conversely, a team might have a perfect balance of roles but lack the specific expertise to execute them. This dilemma, known in research as the balance between individual proficiency and team congruency, has long been a subject of study in organizational science, yet it has remained difficult to measure or predict before a team even begins its work.

A new study published in a scientific journal tackles this problem by looking at one of the largest, most structured environments where temporary teams form every day: the world of competitive online gaming. Researchers focused on League of Legends, a popular game where two teams of five strangers are matched together to compete. In this digital arena, every match is a fresh start. The players are assigned specific roles, such as a defender, a fighter, or a support character, and they must choose from a vast roster of playable characters to fill those roles. The game provider uses a sophisticated system to ensure that the overall skill levels of the two opposing teams are roughly equal, meaning that the outcome of a match is rarely decided by one team simply being "better" than the other. Instead, the winner is often determined by how well the five individuals on each side fit their specific assignments and how well their choices complement one another.

To understand what makes a winning team, the researchers analyzed a massive dataset containing records from approximately 320 million matches involving 26 million unique participants. This scale allowed them to look at the history of every player, tracking which characters they had played, what roles they had filled, and whether they had won or lost in the past. The goal was to build a computer system that could look at a team just before a match began—after the players had been assigned their roles and had selected their characters, but before the game actually started—and predict which team would win. The researchers wanted to know if they could identify a winning composition by understanding two things simultaneously: how well each player matched their current assignment based on their past experience, and how well the five players on a team worked together as a unit.

The team developed a new type of artificial intelligence framework to solve this. Instead of just giving each player a single, static score based on their overall skill, the system looked at each player's history in the context of their current situation. It asked specific questions: How has this player performed with this specific character in this specific role before? How does their past success with this character change depending on the current version of the game? The system then took these individual insights and combined them to see how the five players on a team interacted with each other, and how they matched up against the five players on the opposing team. It paid special attention to the direct matchups, such as how the defender on one team compared to the defender on the other, while also considering the team as a complete, interconnected group.

The results showed that this new approach was significantly better at predicting winners than any previous method. The researchers compared their system to older techniques that simply added up player skill ratings, or methods that looked at team composition without considering individual history. The new system, which learned from the vast history of player choices and outcomes, achieved a much higher accuracy. It proved that knowing a player's general skill level was not enough; the system needed to understand the specific fit between a player and their current role, and how that fit interacted with the rest of the team.

The study also revealed that the two factors—individual fit and team balance—were not just separate ingredients that could be mixed in any order. They were deeply interconnected. When the researchers looked at the data, they found that a team with highly skilled players who were poorly matched to their roles often lost to a team with moderately skilled players who were perfectly balanced. Conversely, a perfectly balanced team could fail if the individual players lacked the specific experience needed for their assigned tasks. The most successful predictions came from a system that treated these two factors as a single, unified problem, learning how the value of a player's past experience changed depending on the team they were in and the opponent they faced.

By analyzing the internal workings of their system, the researchers discovered that it had learned to recognize patterns that humans might miss. For example, the system organized the thousands of different playable characters into groups based on how they functioned in the game, even though it was never told what those functions were. It also learned how the game changed over time, recognizing that a strategy that worked in an older version of the game might not work in a newer one. This ability to adapt to changing contexts and to weigh individual history against team dynamics allowed the system to see the "shape" of a winning team before the first move was made.

The findings suggest that the success of a temporary team is not simply the sum of its parts. It is a complex relationship between who the people are, what they are asked to do, and how those roles fit together in the moment. While the study was conducted in the specific context of a video game, the principles it uncovered offer a new way to think about teamwork in the real world. Whether in a startup, a sports team, or a project group, the research indicates that effective teams are not just collections of talented individuals, but carefully constructed configurations where every member's specific experience aligns with the collective need. The study concludes that to truly understand a team's potential, one must look at the history of its members not in isolation, but in the specific context of the team they have formed.