While we are remarkably capable of generating our own goals, beginning with child's play and continuing into adulthood, we don't yet have computer models for understanding this human ability.
However, a team of New York University scientists has now created a computer model that can represent and generate human-like goals by learning from how people create games. The work , reported in the journal Nature Machine Intelligence, could lead to AI systems that better understand human intentions and more faithfully model and align with our goals. It may also lead to AI systems that can help us design more human-like games.
"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," explains Guy Davidson , the paper's 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."
Despite considerable experimental and computational work on goals and goal-oriented behavior, AI models are still far from capturing the richness of everyday human goals. To address this gap, the paper's authors studied how humans create their own goals, or tasks, in order to potentially illuminate how both are generated.
The researchers began by capturing how humans describe goal-setting actions through a series of online experiments.
They placed participants in a virtual room that contained several objects. The participants were asked to imagine and propose a wide range of playful goals, or games, linked to the room's contents—e.g., bouncing a ball into a bin by first throwing it off a wall or stacking games involving building towers from wooden blocks. The researchers recorded the participants' descriptions of these goals linked to the devised games—nearly 100 games in total. These descriptions formed a dataset of games from which the researchers' model learned.
While human-goal generation may seem limitless, the goals study participants created were guided by a finite number of simple principles of both common sense (goals must be physically plausible) and recombination (new goals are created from shared gameplay elements). For instance, participants created rules in which a ball could realistically be thrown in a bin or bounced off a wall (plausibility) and combined basic throwing elements to create various games (off the wall, onto the bed, throwing from the desk, with or without knocking blocks over, etc., as examples of recombination).
The researchers then trained the AI model to create goal-oriented games using the rules and objectives developed by the human participants. To determine if these AI-created goals aligned with those created by humans, the researchers asked a new group of participants to rate games along several attributes, such as fun, creativity, and difficulty. Participants rated both human-generated and AI-produced games, as in the example below:
Human-created game:
Gameplay: throw a ball so that it touches a wall and then either catch it or touch it
Scoring: you get 1 point for each time you successfully throw the ball, it touches a wall, and you are either holding it again or touching it after its flight
AI-created game:
Gameplay: throw dodgeballs so that they land and come to rest on the top shelf; the game ends after 30 seconds
Scoring: you get 1 point for each dodgeball that is resting on the top shelf at the end of the game
Overall, the human participants gave similar ratings to human-created games and those generated by the AI model. These results indicate that the model successfully captured the ways humans develop new goals and generated its own playful goals that were indistinguishable from human-created ones.
This research helps further our understanding of how we form goals, and how these goals can be represented to computers. It can also help us create systems that aid in designing games and other playful activities.
The paper's other authors are Graham Todd , an NYU doctoral student, Julian Togelius , an associate professor at NYU's Tandon School of Engineering, Todd M. Gureckis , a professor in NYU's Department of Psychology, and Brenden M. Lake , an associate professor in NYU's Center for Data Science and Department of Psychology.
The research was supported by grants from the National Science Foundation (1922658, BCS 2121102).
