Researchers from MIT and other institutions have discovered a phenomenon called the "indoor training effect, " which suggests that training artificial intelligence (AI) agents in less noisy environments can lead to better performance in more unpredictable settings. Traditionally, it was thought that training environments should closely mimic the environments where agents would be deployed. However, this study indicates that when AI agents are trained in stable, noise-free environments, they may outperform those trained in more complex, noisy scenarios when tested. The researchers, led by Serena Bono, explored this effect by training AI to play modified Atari games with added unpredictability.

They found that agents trained in these clean environments performed better overall, supporting the notion that this effect is a broader property of reinforcement learning. Their work challenges conventional wisdom, suggesting that creating simulated training scenarios intentionally designed to minimize noise might be more beneficial. The study involved enhancing exploration strategies of agents, with the understanding that agents trained in less noisy settings learn game rules more easily compared to those exposed to chaos. Moving forward, the team plans to investigate how the indoor training effect can be applied to more complex learning environments and other AI applications, such as natural language processing and computer vision. This research is set to be presented at the Association for the Advancement of Artificial Intelligence Conference.