Automate the optimization loop

Cotinuous learning: analyze, compute, update

4 Layers of Optimization

Changing the way you manage a Free-to-Play game

Target Optimization

Who sees what - Gondola chooses the best offer for a player.

Example: Determine if a new user should see the “Starter Pack” for $0.99 or the “Intro Offer” for $2.99 as the first offer in the game. Note: Gondola does not change $$ or virtual currency amounts for offers, but rather decides which offer from a group of offers to show to a player.

Store Optimization

For in-game stores, Gondola chooses the most relevant items to show to a player.

Example: Your game offers 9 different “Gem Packs” at price points between $2.99 and $199.99. Gondola will compose the right line up of 5 offers in that range for every player cohort.

Currency Optimization

Gondola determines virtual currency quantities for offers and promotions.

Example: For a “Christmas Pack” that costs $9.99, Gondola determines an optimized amount of gold and an optimized amount of coins based on an optimization range set by the product manager.

Video Ad Optimization

Gondola determines the quantity of virtual currency that a player receives in exchange for watching a rewarded video ad.

Example: Watching a 30-second video ad rewards the player with an amount of gems based on an optimization range set by the product manager.

Clustering & Customization

Individual players are assigned to cluster based on a multitude of dimensions. Players move between clusters as they mature in the game. The boundaries of each cluster are constantly reevaluated by Gondola’s machine learning algorithm. The result is that every player is presented with the offers that are most relevant to him.

Improving Games with Multi-Armed Bandit Optimization

Gondola uses a Multi-Armed Bandit (MAB) framework for decision making. As opposed to common A/B testing, where a limited period of pure exploration with rigid traffic allocation to different variants is followed by a period of pure exploitation of the “winning” variants, MAB tests are adaptive. It combines exploration with exploitation (“learn and earn”) by gradually moving traffic towards winning variations, while continuing to test other variants for different player segments.

Gondola’s MAB continuously switches between the learning phase (used to learn how different offers and rewards perform) and the earning phase (using the best offers to maximize performance). The approach works particularly well for free-to-play (F2P) games, since these games are subject to continuous change, requiring continous optimization and testing. As a consequence, our MAB algorithm never declares a winning variant and never phases out other variants completely, but rather takes an adaptive approach to an ever changing game environment. All 4 Gondola Optimization Modules use this technology.

Gondola and the Player Community

Game developers with very active player communities and message boards ask us what the impact of Gondola’s optimization on their community will be. Generally, community backlash is not an issue when using Gondola. However, some games require a more conservative approach, and in these situations we recommend to focus exclusively on Target Optimizations (“who sees what”) and clearly communicate to your players that they are receiving exclusive offers specifically chosen just for them.