Increasing generality in machine learning through procedural content generation

Abstract

Procedural content generation (PCG) refers to the practice of generating game content, such as levels, quests or characters, algorithmically. Motivated by the need to make games replayable, as well as to reduce authoring burden and enable particular aesthetics, many PCG methods have been devised. At the same time that researchers are adapting methods from machine learning (ML) to PCG problems, the ML community has become more interested in PCG-inspired methods. One reason for this development is that ML algorithms often only work for a particular version of a particular task with particular initial parameters. In response, researchers have begun exploring randomization of problem parameters to counteract such overfitting and to allow trained policies to more easily transfer from one environment to another, such as from a simulated robot to a robot in the real world. Here we review existing work on PCG, its overlap with current efforts in ML, and promising new research directions such as procedurally generated learning environments. Although originating in games, we believe PCG algorithms are critical to creating more general machine intelligence. This review covers the history of procedural content generation (PCG) approaches for video games, and how these approaches are now used to generate training data and environments for machine learning models. The authors then discuss how PCG may be crucial for training agents which generalise well.