How to Train Your AI: Researchers Teach AI How to Move Around Fantasy Worlds

Artificial intelligence (AI) tools have grown increasingly complex over the years, being capable of handling a variety of tasks. Now, a research team trained its AI agents on how to act and communicate in a fantasy world.

One of the subsets that have been widely used in AI programming is natural language processing (NLP), which covers interactions between computers and the human language. Previous studies have trained goal-oriented AI agents to play videogames, such as the racing game Gran Turismo. Meanwhile, other studies have integrated NLP techniques to generate more lifelike discourses and intelligent responses. However, NLP techniques and goal-oriented agents have rarely been combined into a single application.

A research team from Georgia Institute of Technology and Facebook AI have collaborated in an attempt to equip goal-driven agents with NLP. It will result in AI players who can speak with other characters and complete quests and tasks within the fantasy game.

Words and Actions Consistent With Overall Goals

Researchers submitted their report, pre-published on arXiv, which illustrates that the combination of goal-oriented agents and NLP strategies result in game players that can communicate and act in a manner consistent with the intended overall goals.

"Agents that communicate with humans and other agents in pursuit of a goal are still quite primitive," said Prithviraj Ammanabrolu, one of the authors of the study from Georgia Tech, in an interview with TechXplore. He added that the study revolves on the premise that most NLP tasks and datasets are static, ignoring a number of previous studies that suggest the necessity of interactivity and language grounding for more effective language learning.

A commonly-used strategy for training AI models is exposure to interactive simulated environments. Interactive narrative games, commonly referred to as text adventures, serve the study in training both goal-driven AI agents and its NLP component, thanks to the games' verbal interactions.

Using Facebook AI's LIGHT

"Interactive narrative games are simulations in which an agent interacts with the world purely through natural language," Ammanabrolu added. In line with this, the ParlAI team at Facebook's AI Research developed LIGHT - Learning in Interactive Games with Humans and Text. LIGHT is a "large-scale fantasy text adventure game" that also doubles as a public research platform for developers of AI models, offering interactions with other AI models or with human users.

LIGHT offers a platform for study and development of grounded dialogue, offering a number of fantasy worlds - each with its own characters, locations, backgrounds, and objects. ParlAI explains that grounding on the local environment allows AI models to progress and generate better predictions for both behavior and dialogue.

Researchers generated a dataset of all quests that can be assigned to in-game characters, which they termed LIGHT-Quests. These goals were crowdsourced and used to provide the agent varying motivations. They then asked other people to play LIGHT and asked insights on how the agents played - how they acted and talked in the game world - as they worked towards their quests.

To illustrate these goals, Ammanabrolu cited a dragon player. In LIGHT, short-term motivation might have something to do with the recovery of a golden egg and seeking punishment for the thief. Long-term goals include, for example, amassing riches to create a large treasure hoard.

Check out more news and information on Artificial Intelligence in Science Times.