While playing a game like an RPG, there will be a lot of gamers who are disappointed that the non-player characters in the village continue to do the same or say the same thing. Through a recent study conducted in collaboration with Georgia Tech and FAIR (Facebook AI Research), we have unveiled a fantasy game AI that communicates and acts with a goal.
In recent years, in the field of AI technology, research on NLP in natural language processing is rapidly developing. In addition, the AI technology that manipulates the characters in the game has also achieved an overwhelmingly advanced AI agent against the top human players.
However, NLP technology and goal-oriented agents that allow game characters to act autonomously have been studied only for individual entities, not several. The Georgia Institute of Technology research team thought that the introduction of NLP technology into goal-oriented agents could develop AI that can communicate and act with other characters, and began research using FAIR’s AI research text adventure game LIGHT.
The research team first created a dataset called LIGHT-Quests by receiving quests that summarized short-, medium- and long-term goals to characters for voluntary collaborators collected online. In Light-Quest, characters such as dragon and knight appear, and the dragon recovers the golden egg stolen by the knight, collects the recovered golden egg in his own treasury, and sets the three-step objective of creating a treasure mountain that no dragon has ever made. Composed with. Light-quest contains 7,486 such quests.
In addition, according to the quest, it was conducted through dialogue between the human player and the AI and collected as training data. There are 2,672 such lines in the Lite-Quest.
The research team created a dataset called ATOMIC-LIGHT that summarizes common sense in the fantasy world, such as that light-quests collect human and AI behavioral and dialogue data, while swords are weapons and kings are royalty, so you should treat them with respect. made. Then, through reinforcement learning, AI was trained through light-quest and atomic-light datasets.
The AI agent developed in this way is said to have achieved high scores as a result of evaluating a statement or action made in a fantasy game through a goal achievement rate and a natural evaluation by humans. For this study, the research team succeeded in creating AI that shows natural behavior using NLP and goal-oriented agents. Reinforcement learning is a way to solve goal-oriented problems, but there was no attempt to combine them with NLP. He said he plans to find a better way to act or talk. Related information can be found here .