Why don't machine learning research agents overfit?

Machine learning research agents, such as AlphaGo and Libratus, have achieved impressive results in various games and domains without overfitting. Researchers have investigated the reasons behind this phenomenon. One reason is that these agents are designed to learn from a large number of training examples, which helps to reduce overfitting. Additionally, many of these agents use techniques such as Bayesian neural networks and regularization to prevent overfitting. Another reason is that these agents are often trained on a variety of tasks and domains, which helps to prevent overfitting by providing a more generalizable learning experience. Furthermore, some researchers have found that the evaluation metrics used in these tasks often prioritize exploration over exploitation, which can also help to prevent overfitting.