An artificial intelligence is incapable of original thought -- at least for now. What it can do, however, is collate information in new ways, resulting in a creation that seems original.
This is how Scheherazade, an artificial intelligence developed by researchers at the Georgia Institute of Technology, works. Named after the storytelling heroine of 1001 Arabian Nights fame, Scheherazade spins interactive fiction of the choose-your-own-adventure variety, with a huge number of branching paths the reader can follow.
This is not the first time an AI has developed game scenarios where player choices determine the outcome; the difference is that Scheherazade uses crowdsourcing to remove the limitations on the number of scenarios that can develop.
Plot points are collected from human users on crowdsourcing platform Amazon Mechanical Turk. The AI then selects and places these plot points in a specific order by determining the basic events for a given situation, and identifying typical event orderings. For example, if John and Sally meet, then go to the movies, they cannot enter the cinema before they have driven there.
"Our open interactive narrative system learns genre models from crowdsourced example stories so that the player can perform different actions and still receive a coherent story experience," said lead investigator Mark Riedl, an associate professor of interactive computing at Georgia Tech.
"When enough data is available and that data sufficiently covers all aspects of the game experience, the system was able to meet or come close to meeting human performance in creating a playable story."
The team tested the AI on two story scenarios, a bank robbery and a movie date, pitting its creations against sets of stories by a human programmed generator and a random generator.
Human test subjects were then divided into three groups. One group read the Scheherazade stories, the second group read the human generator stories, and a third group read the random generator stories.
The groups were tasked with looking for errors, or scenes placed out of sequence. They also had to rate the stories based on their own enjoyment, and how coherent they found them.
The poor old random generator performed poorly in both scenarios, but Scheherazade did almost as well as the human generator. For the bank robbery story, players reported a median of three errors for both the human and AI, while the random generator had a median of 12.5 errors. For the movie date story, the human generator had median of three errors reported, the AI had five, and the random generator had 15.
Additionally, the groups reported similar levels of satisfaction with both the human and Scheherazade stories, indicating they were more or less equally coherent, enjoyable and engaging.
The next step in the research is to try and make the stories generated by Scheherazade more creative and entertaining. Because the system currently only has a basic understanding of stories, it's unable to craft anything more creative than generic experiences. By crowdsourcing more unusual scenarios and resolutions, the team hopes that Scheherazade could craft more unusual stories.
A drama manager could also be used to learn player preferences, and procedurally generate story choices in real-time that are specifically tailored to the individual reader.
Human authors don't have anything to worry about quite yet, though.
"At this point, human-authored interactive narrative still remains the most cost-effective means of generating an interactive narrative experience," the paper concludes.
"However, open interactive narrative shows promise in reducing authorial burden in the near future. Scheherazade-IF and the lessons we learned in creating and evaluating it serve as a first step in creating human-quality interactive narrative with almost no human authoring required."