GVU Center at Georgia Tech | A new Georgia Tech artificial intelligence system develops interactive stories through crowdsourced data for more robust fiction. Here (in a simplified example), the AI replicates a typical first date to the movies (user choices are in red), complete with loud theater talkers and the arm-over-shoulder movie move.

Georgia Institute of Technology researchers have developed a new artificially intelligent system that crowdsources plots for interactive stories, which are popular in video games and let players choose different branching story options.

“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,” says Mark Riedl, lead investigator and associate professor of interactive computing at Georgia Tech.

With potentially limitless crowdsourced plot points, the system could allow for more creative stories and an easier method for interactive narrative generation. For example, imagine a Star Wars game using online fan fiction, generating paths for a player to take.

Current AI models for games have a limited number of scenarios, no matter what a player chooses. They depend on a dataset already programmed into a model by experts.

Near human-level authoring

A test* of the AI system, called Scheherazade IF (Interactive Fiction) — a reference to the fabled Arabic queen and storyteller — showed that it can achieve near human-level authoring, the researchers claim.

“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,” says Riedl.

The creators say that they are seeking to inject more creative scenarios into the system. Right now, the AI plays it safe with the crowdsourced content, producing what one might expect in different genres. But opportunities exist to train Scheherazade (just like its namesake implies) to surprise and immerse those in future interactive experiences.

The impact of this research can support not only online storytelling for entertainment, but also digital storytelling used in online course education or corporate training.

* The researchers evaluated the AI system by measuring the number of “commonsense” errors (e.g. scenes out of sequence) found by players, as well as players’ subjective experiences for things such as enjoyment and coherence of story.

Three test groups played through two interactive stories — a bank robbery and a date to the movies — to measure performance of three narrative generators: the AI story generator, a human-programmed generator, or a random story generator.

For the bank robbery story, the AI system performed identically to the human-programmed generator in terms of errors reported by players, with a median of three each. The random generator produced a median of 12.5 errors reported.

For the movie date scenario, the median values of errors reported were three (human), five (AI) and 15 (random). This shows the AI system performing at 83.3 percent of the human-programmed generator.

As for the play experience itself, the human and AI generators compared favorably for coherence, player involvement, enjoyment and story recognition.

---

Abstract of Crowdsourcing Open Interactive Narrative

Interactive narrative is a form of digital interactive experience in which users influence a dramatic storyline through their actions. Artificial intelligence approaches to interactive narrative use a domain model to determine how the narrative should unfold based on user actions. However, domain models for interactive narrative require artificial intelligence and knowledge representation expertise. We present open interactive narrative, the problem of generating an interactive narrative experience about any possible topic. We present an open interactive narrative system— Scherazade IF—that learns a domain model from crowdsourced example stories so that the player can perform different actions and still receive a coherent story experience. We report on an evaluation of our system showing near-human level authoring