Learned Game Engine

A learned game engine is a system where the core functions of a traditional game engine — physics simulation, game logic, and visual rendering — are replaced by neural networks trained directly on data rather than hand-authored rules.

In a learned game engine, a world model learns to predict state transitions from player actions, essentially acquiring the gameplay dynamics that would normally be written as explicit code. A separate rendering model then generates visual output conditioned on the evolving game state.

Both components are entirely learned systems that extract patterns from training data, rather than implementing known algorithms. This allows learned game engines to generalize beyond their training — generating novel levels and scenarios that remain consistent with the dynamics of the source games.

The concept represents a shift from explicit programming of game behavior toward emergent game intelligence learned from observation and interaction.