Sony has filed a patent that illustrates how a game's non-playable characters (including enemies and bosses) could use a machine learning (ML) model to dynamically adjust to a player's behaviors, which could allow some single-player games to gain levels of depth and difficulty that would normally require a human enemy player.

For the game genres that would be impacted, virtually any combat-centric single-player game could be enhanced by a system like this. The game itself doesn't need to have complex systems: the fundamentals of PVP play in the fighting game genre, for example, revolve around "footsies," or what players of other games would call "positioning" or "spacing." An ML AI model could be transformative for single-player games, even if its functions were as simple as controlling enemy spacing and attack timing.

Sony PlayStation ML-enabled enemy AI patent illustration. Image: Sony (via Patentlyze)

That isn't to say this is completely good news. Sometimes, being able to "solve" a single-player game down to the minutiae of exact enemy behaviors is what makes a game fun and replayable. But then, many speedrunners also enjoy "randomizers" and other mods that exist explicitly to diminish the importance of memorization. Design-wise, a feature like the one outlined in the patent would certainly make a game more complex, but not necessarily "more fun".

If a complex single-player action or adventure game could be designed to learn how to counteract a player's strategy, however, the theoretical replay value of such a title could be near-infinite. A number of complex, niche subgenres of action games (Ninja Gaiden, Dark Souls, Devil May Cry, etc.) have already been played for decades without adaptable AI, though.

The new Sony patent unveiled by Patentlyze does sound interesting, but there are some obvious potential downsides. The biggest problem is that we may see a repeat of what happened with the Nemesis system from Middle-Earth: Shadow of Mordor, where developer Monolith and publisher Warner Bros, enforced a patent preventing similar systems from being implemented in other games. A feature like adaptive enemy AI sounds awesome, but a little depressing if it's locked to digital-only PlayStation 6 games.

A machine learning model isn't actually needed to make adaptive enemies. In fighting games like Tekken 8 or Street Fighter 6, incredibly detailed player Ghosts learn common player behaviors and responses already. Single-player content in those and other games have also used features like input reading, where the enemy has perfect reflexes to everything you do because it knows as soon as the button is pressed.

Still, using machine learning to adapt and customize gameplay to a specific gamer's style and capabilities could spur some interesting innovations and make tomorrow's games even more realistic and immersive.