Story
Cinder: A fast and fair matchmaking system
Key takeaway
Researchers developed Cinder, a new matchmaking system that can quickly pair players in online games into fair matches, which helps keep players engaged and satisfied with the gameplay experience.
Quick Explainer
Cinder is a two-stage matchmaking system designed to create fair matches in online games, even when players form pre-made lobbies with widely varying skill levels. The first stage quickly filters out incompatible lobby pairings by comparing the central skill clusters. The second stage calculates a comprehensive "fairness score" by mapping player skills to non-linear "buckets" and quantifying the distance between the skill distributions of the lobbies. This advanced fairness metric goes beyond relying on simple statistical measures like average skill, which can fail to capture the full complexity of skill disparities within a lobby.
Deep Dive
Technical Deep Dive: Cinder: A Fast and Fair Matchmaking System
Overview
Cinder is a novel two-stage matchmaking system designed to create fair matches in online games, even for lobbies with high skill variance between players. It addresses the limitations of existing matchmaking approaches that rely on simple statistical measures like mean or median skill rank, which can fail to accurately represent the skill distribution in a lobby.
Problem & Context
In online games, the matchmaking system is crucial for pairing players into fair and engaging matches. This becomes challenging when players form pre-made "lobbies" with friends of widely varying skill levels. Traditional matchmaking based on average skill can create unfair experiences, with low-skilled players overwhelmed and high-skilled players bored.
Methodology
Cinder's two-stage approach:
- Fast Preliminary Filtering: A computationally inexpensive check to quickly discard incompatible lobby pairings by comparing the "non-outlier range" (central skill cluster) of each lobby using the Ruzicka similarity index.
- Accurate Fairness Quantification: A more precise evaluation that considers the full skill distribution of all players in each lobby. It maps player skills to non-linear "skill buckets" and calculates the Wasserstein (or "Sanction") distance between the lobbies' bucket distributions to generate a single fairness score.
Results
Simulations of 140 million random lobby pairings showed the Sanction Score follows a right-skewed distribution, with the majority of pairings clustered around a moderate level of "unfairness" and increasingly rare instances of extremely poor matches. This empirical distribution can be used to set an optimal fairness threshold, balancing match quality and queue time.
Limitations & Uncertainties
The paper does not report results from live deployments of Cinder, so its real-world performance and impact on player satisfaction remain to be seen. Further research could explore integrating additional factors like player wait time or role preferences into the fairness calculation.
What Comes Next
Future work could involve:
- Testing Cinder in live game environments to measure its impact
- Exploring integration of other factors (wait time, role preferences, etc.) into the fairness score
- Optimizing the non-linear bucket distribution based on a game's specific player-rank demographics
