Dungeon Master
The artificial intelligence that completes quests in a text-based adventure game by talking to characters has learned to not only act on its own, but also force others to do something. Such a system is a step towards creating machines that can use natural language as a way to achieve their goals.
Natural language model processing algorithms like GPT-3 are great at mimicking human-written sentences, churning out stories, fake blogs, and posts. But this fruitful product doesn't make much sense other than creating the text itself. When people use language, they use it as a tool: our words convince, command and manipulate; they make people laugh and cry.
To create AI that makes good use of words, researchers at Georgia Institute of Technology in Atlanta and Facebook AI Research have combined natural language processing with learning, which learns machine learning models to achieve their intended goals.
How to communicate with a dragon?
The researchers trained their system in a text-based multiplayer game called LIGHT, developed by Facebook last year to study communication between humans and AI players. The game takes place in a themed fantasy world filled with thousands of crowdsourced objects, characters and locations that are described and interact with on-screen text. Players (human or computer) act by choosing commands such as hug the wizard, kill the dragon, or take off the hat. They can also talk to characters controlled by chatbots.
Artificial intelligence in this game got the role of a dragon, which received certain missions, for example, the accumulation of gold. To be successful in completing tasks, he had to communicate with other AI agents or real gamers, simply entering certain commands, like in any other text-based adventure.
The results were a bit bizarre, with the dragon issuing meaningless threats to force the characters to fulfill its wishes. However, according to the team's research, AI still achieved its goal due to the fact that it began to understand the individual characteristics of communication between different real characters.
. , , , , . .
To explain to their AI the reason it does something, the researchers added about 7,500 crowdsourced quests not included in the original version of LIGHT. They also created a knowledge graph (a database of subject-relationship-object relationships) that gave the AI robust information about the game world and the relationships between its characters. For example, the principle that a merchant will only trust a guard if they are friends. The game introduced actions such as "Go to the mountains" and "Eat a knight", which must be completed to complete quests (for example, "Build the largest treasure the dragon has ever found").
Putting it all together, the developers trained the AI to complete quests simply using natural language. To perform any of the actions, a person can either enter the appropriate command, or achieve the same result by talking to other characters. For example, if an AI needs a sword, it can steal it or convince another character to give it back.
An example of a dialogue of a person (gray) performing his mission with AI (blue).
Of course, today this system is just a toy. And the manner in which artificial intelligence communicates with real players may seem quite straightforward: at some point, when it needs a bucket, it simply says: "Give me this bucket, or I'll feed you to my cat!" But mixing natural language processing with learning is an exciting step that can lead not only to better chatbots that can argue and persuade, but also those that have a better understanding of our language-filled world.
List of references:
- To teach an AI to purpose goals, scientists made it play an RPG [Electronic resource]
- How role-playing a dragon can teach an AI to manipulate and persuade // MIT Technology Review [Electronic resource]
- How to motivate your dragon: teaching goal-driven agents to speak and act in fantasy worlds // Official report of the developers [Electronic resource]