Have you ever wondered how computers could learn to set goals like we do? It’s a fascinating challenge because while humans naturally set goals from a young age, capturing this in a computer model has been a tough nut to crack.
But here’s some exciting news: researchers at New York University have made a breakthrough! They’ve developed an AI that learns to set goals just like us by observing how we create games.
This study, published in Nature Machine Intelligence, could be a game-changer in how AI understands human intentions, potentially leading to AI systems that can design games with a personal touch.
Guy Davidson, the study’s lead author, sums it up nicely: “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.” This research offers a new way to understand how we create and think about goals, which could lead to more creative and effective AI systems.
Despite years of research, AI models have struggled to grasp the full depth of human goal-setting. So, what did the researchers do differently? They looked into how we come up with tasks. Through online experiments, participants were asked to propose playful goals in a virtual space filled with objects. Imagine games like bouncing a ball into a bin or stacking blocks. The result? A dataset of nearly 100 games!
This dataset showed that while human goal creation feels limitless, it’s actually guided by simple principles of plausibility and recombination. Participants came up with feasible rules, like throwing a ball into a bin, and mixed elements to create varied games, such as bouncing a ball off a wall.
The team then trained the AI to develop games using these human principles. To see how well the AI did, a new group of participants rated the games on fun and creativity. They checked out both human-created and AI-generated games. For example:
- Human-created game: Throw a ball to touch a wall and catch it.
- AI-created game: Throw dodgeballs to rest on a top shelf within 30 seconds.
The results were fascinating! Ratings showed that AI-generated games were just as good as those made by humans. This indicates that the model was successful in mimicking human goal development. This research not only sheds light on how we form goals but also holds promise for improving AI systems in game design and beyond.
