Machine Learning for Optimal Matchmaking

Explore TrueMatch, a revolutionary matchmaking approach presented by 343 Industries' Josh Menke in this 2020 GDC Virtual Talk. Discover how machine learning can be leveraged to automatically optimize desired metrics in real-time, allowing developers to express the value of each metric more intuitively. Delve into the intricacies of matchmaking, machine learning, and their intersection in game development. Learn about the current state of skill-based matchmaking and real-time techniques. Understand the limitations of conventional approaches and how TrueMatch offers a more optimal solution through unified objective functions and utility functions. Examine the components of the TrueMatch algorithm, including population trackers, metric predictors, and optimizers. Gain insights into handling regional matchmaking and free-for-all scenarios. Compare conventional and TrueMatch approaches through practical examples and visualizations. Conclude with simple improvements and key takeaways to enhance your game's matchmaking system.

Intro
What is matchmaking?
What is Machine Learning?
Machine Learning and Matchmaking
State of the Art Today: Skill
Real-time matchmaking
First Configure a Set of Rules
Example Configuration
Matchmaker Receives Requests
Matchmaker Request
Compares Requests
Creates a Match
Rules Apply Globally
Predictability How often the better team wins
Conventional Takeaway
More Optimal Approach
Defining Optimal
High-level Comparison
Unified objective function
As an Equation
Utility Function Use
True Match algorithm
True Match components
Population tracker
Population model
Metric Predictor
Simplest wait time formula • matchable(t) = requests that can match with t
Parameterized wait time formula
Optimizer: How does it make Rules?
Start with Current Rules
Find Optimal Transform Need to rewrite to search all curves
We found it!
True Match Rule Example Scale = 10, remember gap of 1 is OK
Mapped vs. Conventional
True Match Rules
What about Regions?
Conventional Region Approach
True Match Approach
FFA example
Skill gap variation
Wait time variation
True Match FFA Take-away
True Match Takeaway
Simple Improvements Matchmake on scaled Skill Percentiles
Thank you! Questions?