King

Objective

Increase engagement and conversion on in-game offers by serving each player the most relevant bundle, while producing and testing Candy Crush levels at scale.

The deployment

King recommends personalized bundles in Candy Crush Saga with a machine learning system that combines supervised and unsupervised learning and a so-called scale-invariant prediction approach, backed by player-level attention-based models. A paper authored by King teams reports a 30 percent increase in engagement on click rate and more than 40 percent on take rate, with an explicit framing of deployment at the scale of millions of users, of model maintenance and monitoring, and of degenerate feedback loops to watch. In parallel, AI feeds game production: according to the AI Labs director, nearly every Candy Crush level is now designed with the help of AI, with bots playing each level thousands of times to measure difficulty before designers validate. King has also built BAIT, a testing bot that traverses the game through screen capture and interface recognition to detect missing textures, absent text, and rendering artifacts.

Results Proof B

Technical paper authored by King teams, with figures on click rate and take rate of a system deployed at the scale of millions of players, corroborated by specialized press and an engineering article on automated testing. Not a financial document, so B.

1. 1
   Automated level testing AI (test bots)

   AI bots play each level thousands of times to measure difficulty and friction; BAIT detects visual bugs.

2. 2
   Design validation designers (humans)

   Designers read the AI metrics, annotate good and bad examples, and decide whether to adjust.

3. 3
   Offer prediction AI (recommendation system)

   The attention-based model predicts, per player, the bundle with the highest conversion propensity.

4. 4
   Display and tracking data / LiveOps team

   The store serves the offer; click rate, take rate, diversity, and feedback loops are monitored.