NYC Gazette

A team of scientists from New York University has developed a computer model capable of generating human-like goals by learning from the way people create games. This research, published in the journal Nature Machine Intelligence, aims to enhance AI systems' understanding of human intentions and align them more closely with human goals.

"While goals are fundamental to human behavior, we know very little about how people represent and come up with them—and lack models that capture the richness and creativity of human-generated goals," said Guy Davidson, the lead author and an NYU doctoral student. "Our research provides a new framework for understanding how people create and represent goals, which could help develop more creative, original, and effective AI systems."

The researchers explored how humans devise their own tasks or goals to potentially shed light on this process. The study involved online experiments where participants were placed in a virtual room containing various objects. They were asked to imagine and propose playful goals or games related to these objects. Examples included bouncing a ball into a bin by first throwing it off a wall or stacking wooden blocks into towers. These activities resulted in nearly 100 different game descriptions forming a dataset from which the model learned.

Despite the seemingly limitless nature of human goal generation, participants' creations adhered to simple principles such as physical plausibility and recombination. For instance, rules were created where balls could be thrown realistically into bins or bounced off walls.

The AI model was then trained using these principles to create goal-oriented games. To assess whether AI-generated goals matched those created by humans, another group of participants rated both sets of games based on attributes like fun, creativity, and difficulty.

Overall ratings indicated that the AI model successfully captured how humans develop new goals and generated playful ones indistinguishable from those created by humans.

This study contributes to our understanding of goal formation and representation for computers while aiding in designing games and other playful activities.

Other authors include Graham Todd (NYU doctoral student), Julian Togelius (associate professor at NYU’s Tandon School of Engineering), Todd M. Gureckis (professor in NYU’s Department of Psychology), and Brenden M. Lake (associate professor at NYU’s Center for Data Science and Department of Psychology). The research received support through grants from the National Science Foundation (1922658, BCS 2121102).