Sophistication of Human Adaptive Probability Learning in Dynamic Environments

Wai Ying Chung, Florent Meyniel

Abstract: Maintaining accurate beliefs in a changing and noisy environment is a challenging computational problem. Previous studies have shown that humans adapt their learning dynamically, especially in the face of change. This conclusion is mostly supported in the context of magnitude learning (e.g., tracking a reward amount, an object position), and currently remains more uncertain in the case of probability learning (e.g., tracking the probability of an event occurring). Here, we initiate an open benchmarking approach to uncover the computations humans use for probability learning. We compared a wide range of models—including optimal Bayesian models, suboptimal variants, and simple prediction error-based update rules, using several datasets in which participants provided trial-by-trial probability estimates. Bayesian inference often outperformed simple prediction error-based models, despite being more computationally demanding and often considered less biologically plausible. Furthermore, inference strategies appear to depend on environmental volatility: under moderate volatility, an optimal Bayesian model best explains behavior, whereas in more stable environments, a simpler Bayesian approximation is better. These results so far highlight the sophistication of human adaptive learning for probability and suggest that humans can adapt their inference strategies based on environmental context. We invite others to contribute models and datasets to this benchmark to refine these conclusions.

Submission Number: 40