AI Machine Learning & Data Science Nature Language Tech Research

Facebook AI | Text-Based Generation of Interactive Worlds

Researchers recently proposed a new machine learning method for worldbuilding based on content from LIGHT, a research environment open-sourced by Facebook comprising crowd-sourced game locations, characters, and objects, etc.

Central to the appeal of many computer games is their complex environments. As a significant component of contemporary game design and storytelling, “worldbuilding” is the art of creating engaging locations with interesting characters, objects backgrounds and other details as the basis for one or more game tasks or stories. In text-based games — where the atmosphere and action is conveyed with words rather than pictures — worldbuilding can present an even more challenging task for AI models.

In an attempt to make the relationship between such game elements more natural, researchers from Facebook AI Research, French computer science research lab Loria, and University College London recently proposed a new machine learning method for worldbuilding based on content from LIGHT, a research environment open-sourced by Facebook comprising crowd-sourced game locations, characters, and objects, etc.

The researchers propose a neural network based solution that can automatically structure and arrange locations, characters, and objects into a single whole and coherent game environment. The process begins by initializing an empty grid, where each square can host different possible locations (nameand description). One of these locations is randomly selected as the starting point, then the neural network model will proceed to detect and fill neighboring locations — a process made more natural via crowd-sourced example neighbors data. As the model fills the locations with appropriate characters and objects, it also decides whether and where to place objects in containers, and predicts the potential and compatibility of new object combinations with different locations.

In addition to building environments based on existing game elements, the model can also generate and develop entirely new game elements and content.

Screen Shot 2019-12-23 at 12.56.22 PM.png
Sample constructed game world. Models arrange locations, then populate them with characters and objects.

In comparisons with other machine learning based worldbuilding algorithms, human evaluators ranked game environments generated by the new method as more “cohesive, diverse and interesting.” The researchers say their models can also assist players in designing their own game environments by providing suggestions on what elements to fill into different locations.

The paper Generating Interactive Worlds with Text paper has been accepted by AAAI 2020 and is available on arXiv.


Author: Herin Zhao | Editor: Michael Sarazen

2 comments on “Facebook AI | Text-Based Generation of Interactive Worlds

  1. Pingback: Facebook AI | Text-Based Generation of Interactive Worlds - PC Solution

  2. Thanks for your efforts.

Leave a Reply

Your email address will not be published. Required fields are marked *

%d