Collective Intelligence Outperforms Individual Talent: A Case Study in League of Legends

TL;DR

The paper investigates whether collective intelligence arising from cooperative behavior can outperform talented individuals in a MOBA game, League of Legends. It employs a graph-spectral framework, using directed graphs of in-game assists and the novel Effective Graph Resistance as a proxy for cooperation, together with exploratory factor analysis and ML classifiers, on a large-scale LoL dataset from Riot API. The main contributions include validating H1 that collective intelligence beats individual intelligence, showing that highly skilled players exhibit higher collective intelligence, and providing visual analytics (chord diagrams) that link interaction topology to win outcomes. The findings have implications for team formation and governance in complex social-technical systems beyond gaming.

Abstract

Gaming environments are popular testbeds for studying human interactions and behaviors in complex artificial intelligence systems. Particularly, in multiplayer online battle arena (MOBA) games, individuals collaborate in virtual environments of high realism that involves real-time strategic decision-making and trade-offs on resource management, information collection and sharing, team synergy and collective dynamics. This paper explores whether collective intelligence, emerging from cooperative behaviours exhibited by a group of individuals, who are not necessarily skillful but effectively engage in collaborative problem-solving tasks, exceeds individual intelligence observed within skillful individuals. This is shown via a case study in League of Legends, using machine learning algorithms and statistical methods applied to large-scale data collected for the same purpose. By modelling systematically game-specific metrics but also new game-agnostic topological and graph spectra measures of cooperative interactions, we demonstrate compelling insights about the superior performance of collective intelligence.

Figures (7)

• Figure 1: (A) The three lanes, 2 jungles and their subdivisions. Small red and blue circles denote tower structures for the respective teams. The darker, inner-circle segment in the top-right and bottom-left corners denote each teams’ Nexus; (B) An instance of gameplay on Summoner's Rift where an opponent kill is scored due to positional map-pressure. Blue team player A attempts to kill red team player Y, who tries to flee. Blue player B blocks Y's escape, forcing engagement and leading to A's kill. Although B does not actively assist, B’s positioning creates map pressure resulting in Y’s death. B is awarded a map-pressure assist if it is positioned within a specified radius during the kill. (C) Details of an in-game champion;
• Figure 2: Overall architectural framework for the evaluation of collective intelligence hypotheses within League of Legends gameplay environments. The architecture comprises two main components: (i) computation of individual and team metrics, and (ii) analysis of winning outcomes using exploratory data analysis with classification and clustering methods on such individual and collective team-level data.
• Figure 3: Directed graph network for a sample team characterizing the flow of interactions between players. Each graph has 5 nodes corresponding to player roles and weighted, directed edges signify the frequency and flow of all forms of assistance from one player to another. Average values of team performance and computed graph metrics is shown.
• Figure 4: Feature importance values for individual and team collective performance factors across different classification algorithms. Across all classification models, at the individual level, the factor which represents acquisition of resources for a player is more crucial in winning the game rather than sharing. However, at the collective team level, cooperative behaviors such as resource sharing takes precedence over non-cooperative behaviors such as centralization of resources.
• Figure 5: Frequency of wins for cooperative teams having a majority of average-to-low skilled individuals and non-cooperative teams having a majority of low-to-average skilled individuals when they face each other.