Abstract
In large-scale game development, artificial intelligence is often siloed within individual titles, leading to redundant systems and limited cross-project utility. This paper, presented at IAAI-18 by researchers at Electronic Arts (EA), details a centralized AI platform designed for "horizontal scaling." By treating AI as a service rather than a bespoke game-playing agent, EA has successfully deployed machine learning solutions across a diverse portfolio of AAA and mobile titles, improving player retention, matchmaking, and operational efficiency. This breadth-first strategy demonstrates how a unified framework can overcome the technical and organizational hurdles of deploying AI at scale.
Introduction
Large gaming organizations like EA face a unique challenge: managing a decentralized ecosystem of studios, each with its own tech stack. When new AI trends emerge, teams often race to build bespoke, redundant solutions, leading to "tool proliferation." The authors argue that the primary hurdle in industrial AI is not model accuracy, but the ability to deliver AI artifacts consistently across the entire company. To address this, they developed a framework that shifts the focus from "AI as a game-playing agent" to "AI as a service" that optimizes the end-to-end player experience, from the store to the matchmaking lobby.
The Three Axes of Scaling
The framework categorizes AI development along three distinct axes: Wide (data volume), Deep (model complexity), and Broad (number of projects). While the industry often focuses on "Deep" scaling—such as creating superhuman agents for games like Go—EA prioritizes "Broad" scaling. This strategy aims to deliver hundreds of "small" wins across many games rather than a single high-effort "moon-shot" project. By focusing on breadth, the company can achieve a cumulative impact that far outweighs the benefits of a single, isolated breakthrough.
System Architecture
The platform's architecture is built to minimize integration friction and maximize reuse. Central to this is the Agent Store, a repository that treats models as "first-class citizens."
Key components of the architecture include:
- Portable Format for Analytics (PFA): Models are represented as Turing-complete JSON/YAML code. This allows a model to be generated once and run across different game engines without being rewritten.
- Recommendation Engine (Reco): A microservice that uses prioritized rules (written in JRuby or Javascript) to decide actions based on player history and current context.
- Experimentation Microservice: A built-in A/B testing framework that links AI actions directly to telemetry, allowing the team to measure the impact of a model on Key Performance Metrics (KPMs) in real-time.
- Data Warehouse: A massive Hadoop/Hive-based system that curates anonymized player history to feed the models.
Key Results
The horizontal scaling approach has yielded measurable improvements across several domains using internal telemetry data from titles like Battlefield 1 and Bejeweled Stars:
- Dynamic Difficulty Adjustment (DDA): By using churn prediction to tune level difficulty, EA increased games played per day and session time by approximately 10%. A later iteration using dynamic programming boosted this by another 7%.
- Content Recommendation: In Battlefield 1, using the Reco engine to suggest maps and modes increased 7-day retention by 18%.
- Monetization: Implementing a multi-armed bandit mechanism for product recommendations in the Origin store led to a 12% increase in click-through rates.
- Anti-Cheat: The "Bad Actor Detection System" (BADS) has identified and banned hundreds of thousands of fraudulent accounts with a false positive rate of less than 1%.
- Game Balancing: AI agents simulated 1,000 games of The Sims Mobile in an hour, allowing for rapid tuning of game economies that would otherwise take human specialists days.
Conclusion
The success of EA's platform demonstrates that in an industrial context, the system surrounding a model is as important as the model itself. By prioritizing breadth and repeatable processes over one-off solutions, the framework allows for rapid deployment and iteration. The shift from "AI as a thermometer" (measuring state) to "AI as a thermostat" (changing state) has proven to be a powerful driver of player engagement and operational efficiency across a massive software ecosystem.
Summary updated