Microsoft patents machine learning system that predicts optimal times for in-game ads

Microsoft has filed a patent application for an in-game advertising system that could use machine learning to monitor player behaviour and work out when you are least likely to be furious about a sudden ad. The application, titled "Contextually Aware Management of Interactive Software Application Access," was filed in April and published on 3rd September by the United States Patent & Trademark Office.

For players, it is another sign that in-game advertising is heading somewhere more invasive: the patent describes a system that doesn't just slam an ad on screen, but studies what is happening in your game and waits for the calmest possible moment to strike.

How the system would work

In broad strokes, the hypothetical system would pause gameplay to show adverts, but only when the game's context suggests "adverse occurrences are unlikely or not possible to occur" — for instance, when you are in a safe location — and ideally without creating what the filing calls a "negative use experience".

There are plenty of existing ad systems, particularly in free-to-play mobile gaming, but this package arguably points at a new frontier in automated surveillance and monetisation. The full document, via Respawn First, is dense legalese, but the underlying idea is clear: reshape the experience of playing a game around the experience of watching ads.

Earning credits to delay the next ad

One described implementation could award players a visible or invisible "credit" each time they view a promotion, with credits delaying the next advert. You might be able to bank several at once or "view another promoted content to earn additional credits". The system might even tell you roughly how long the next ad will run, or whether watching one now makes sense given your current in-game situation.

Ad timing could be tied directly to gameplay milestones, with the filing citing triggers like "5 rounds of gameplay, completing one level of a video game, completing 3 online matches, defeating a non-playable character (NPC), or completing a project or task".

Reading the game over your shoulder

The real ambition is the system's responsiveness. It could hold off during a "rare event" — a moment of "excitement and/or curiosity" for certain player types. It might detect when you open a save screen or dialogue box, pause the game, or finish a round. It could even notice texture changes showing you've taken damage or swapped gear, and learn to tell apart different art styles for the same object (the authors point out trees in Call of Duty don't look like trees in Fortnite).

Machine learning would likely drive much of this, analysing "video information, audio information [or] game state data" and refining itself over iterations — for example, learning to associate health packs with health bars increasing.

The system might also "learn" to avoid interrupting stream-friendly moments: a spectacular catch in a baseball game, a terrain exploit during a speedrun, a legendary loot drop, or a fighting game player landing a tricky parry.

For now, it remains just a patent application, and not everything filed ever ships. EA, separately, has said it plans to experiment further with in-game advertising billboards — which, next to an AI reading your game in real time, sounds almost quaint.

Source: Rock Paper Shotgun