NYU scientists developed a new framework to better understand how people set goals when trying to win a game, hit a benchmark or achieve some other objective. The researchers aimed to generate effective artificial intelligence models that can mimic human behavior in goal-oriented settings.
First, researchers collected a dataset of human-generated goals by asking participants to describe a scorable game using objects such as cubes, balls and walls. Then, the researchers translated the games into a computer program to create “domain-specific language” — a programming language tailored to a specific goal — that modeled the semantics of human-created games’ rules and scoring methods.
“There’s wonderful work on what the utility of playful goals and games is,” Guy Davidson, lead author of the study and Ph.D. candidate at the NYU Center for Data Science, wrote in a statement to WSN. “But there’s no description of how we create these goals and what sort of representation might we have for them in our mind.”
When the researchers input the DSL model into a second computer program, they found that it was able to create goals that emulate and expand upon human thought processes without exactly replicating the initial input. In an interview with WSN, co-author and Ph.D. student Graham Todd said they evaluated how the objectives in human and computer-generated games compared to assess how accurately the program worked.
The researchers defined effective goals as ones that were physically reasonable and emulated humans’ “intuitive common sense.” For instance, two balls stacked on top of each other wouldn’t be considered a strong goal because it’s physically improbable — although this might not have been spelled out in the human’s initial input.
“We use this function that’s saying, ‘How human-like is this goal?’ That is the measure of quality,” Todd said. “Then you also have some notion of diversity — we want a whole big diverse set of goals, but we want them all to be good.”
Todd said the research has implications for improving AI and machine learning tactics. He said that by creating models that demonstrate how humans form goals, scientists can program computers to perform increasingly complex and relevant functions.
“We want artificial intelligence agents to constantly be learning and trying to do new things,” Todd said. “When they accomplish it, instead of us telling the AI system what to do next, a system which is like a person is able to decide for itself what it wants to do.”
Contact Gabrielle Panelo at [email protected].