AI agents that “self-reflect” perform better in changing environments

In a head-to-head competition between a state-of-the-art AI agent and a mouse, the surprising winner was the mouse. The researchers, Isaac Kauvar and Chris Doyle, placed a mouse in a small empty box and a simulated AI agent in a 3D virtual arena, both with a red ball. The mouse quickly approached and interacted with the ball, while the AI agent didn’t notice it. Inspired by the mouse’s behavior, the researchers developed a new training method called curious replay, which improved the AI agent’s performance. This method encourages AI agents to self-reflect on novel and interesting experiences, leading to better exploration and adaptation. The researchers believe this approach will have wide-ranging applications in AI research.

To top