AI and artificial games create playful games

New research reveals how computers can generate human-like goals through comprehensive learning of programs

The difference between stacking obstacles falling and trying the AI ​​for the same task is always clear: one is driven by real playfulness and creativity, while the other follows programming instructions. However, according to groundbreaking research by scientists at New York University, this month Natural machine intelligence.

In the study, researchers developed a computational model that represents and produces human-like goals by learning how people are creative. The model proved to be so effective that when human evaluators were asked to evaluate the game, they could not reliably distinguish between humans and those produced by AI.

“While goals are the basis of human behavior, we have a lack of models for how people represent and present goals, and we have a lack of models that capture the richness and creativity of humans that generate goals,” explained Guy Davidson, lead author of the paper and a doctoral student at New York University. “Our research provides a new framework for understanding how people create and represent goals that can help develop more creative, primitive and effective AI systems.”

The research team, including Graham Todd, Julian Togelius, Todd M. Gureckis, and Brenden M. Lake, started with an online experiment in which participants were placed in a virtual room containing various objects such as balls, blocks and furniture. Participants were asked to invent single-player games using only objects that exist, thus creating nearly 100 different games.

From children’s games to computer understanding

What makes research particularly noteworthy is how it bridges cognitive science and artificial intelligence. Although most AI models that deal with goals depend on simple parameters such as “reach the target position” or “win the game,” the diversity and creativity of human goals.

The researchers observed that despite seemingly unlimited possibilities, artificially created goals followed a pattern guided by common sense (body rationality) and reorganization (mixing familiar elements in new ways). For example, participants instinctively knew that they could throw balls into a trash can or bounce off a wall, and they combined these basic moves to create various games.

Scientists represent these goals as “a program that generates rewards,” i.e., a consistent operation that evaluates progress and provides feedback. This approach allows them to identify the composition patterns of different games.

The paper states: “Participants demonstrated intuitive common sense.” When creating throwing games, the majority of participants chose balls rather than other objects. Again, for stacking games, they use blocks primarily rather than balls. This intuitive understanding of humans seems obvious, but it is a major challenge for AI systems.

Teach AI to play like a human

The team then built a Target Program Generator (GPG) model that learns to generate new games based on patterns observed in games created by humans. The model includes explicitly designed components to approximate cognitive abilities such as physical common sense and reorganization.

When evaluating the output of the model, the researchers divided AI-generated games into two types: those matching patterns found in games created by humans and those exploring new domains. For example:

Games created by humans:

• Gameplay: Throw a ball, make it touch the wall, then grab it or touch it
• Score: Every time you throw the ball successfully, you will get 1 point

Games created by AI:

• Gameplay: Throw dodge Powers, let them land and rest on the top shelf; the game ends in 30 seconds
• Score: 1 point per dodge ball resting on top shelf at the end of the game

Individual human evaluators rate games created by humans and AI as factors such as entertainment, creativity, and difficulty. It is worth noting that participants generated similar ratings for games created by humans, as well as games produced by AI models matching human patterns, which are equally understandable and enjoyable.

Beyond Gamification

The meaning goes far beyond creating a better game. This research helps promote our understanding of how humans form goals and how they represent them to computers, a crucial challenge in developing AI that truly understands human intentions.

“Understanding how humans create, represent, and reasoning goals is essential to understand human behavior,” the study authors wrote. “People often create novel, trait-based goals that go beyond these common modeling settings.”

The research team suggested that their framework could enhance AI systems’ ability to explore and adapt to new environments. It can also improve how machines interpret human intentions, a key factor in developing AI that is aligned with human values.

Supported by a grant from the National Science Foundation, this study is an important step towards AI systems that not only follow instructions but also understand the creative, interesting goals that drive human behavior from childhood.