A bias–variance trade-off governs individual differences in on-line learning in an unpredictable environment

• Published in: Nature human behaviour Vol. 2; no. 3; pp. 213 - 224
• Main Authors: Glaze, Christopher M., Filipowicz, Alexandre L. S., Kable, Joseph W., Balasubramanian, Vijay, Gold, Joshua I.
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 01.03.2018; Nature Publishing Group
• Subjects: 631/378/116/2396; 631/378/2649/1409; Algorithms; Behavioral Sciences; Bias; Biomedical and Life Sciences; Decision making; Experimental Psychology; Individual differences; Learning environment; Life Sciences; Microeconomics; Neurosciences; Personality and Social Psychology; Research subjects; Sampling; Variability
• ISSN: 2397-3374; 2397-3374
• DOI: 10.1038/s41562-018-0297-4

Decisions often benefit from learned expectations about the sequential structure of the evidence. Here we show that individual differences in this learning process can reflect different implicit assumptions about sequence complexity, leading to performance trade-offs. For a task requiring decisions about dynamic evidence streams, human subjects with more flexible, history-dependent choices (low bias) had greater trial-to-trial choice variability (high variance). In contrast, subjects with more history-independent choices (high bias) were more predictable (low variance). We accounted for these behaviours using models in which assumed complexity was encoded by the size of the hypothesis space over the latent rate of change of the source of evidence. The most parsimonious model used an efficient sampling algorithm in which the range of sampled hypotheses represented an information bottleneck that gave rise to a bias–variance trade-off. This trade-off, which is well known in machine learning, may thus also have broad applicability to human decision-making. Glaze et al. show that individual variability in learning from noisy evidence involves a bias–variance trade-off that is best explained by a model using a sampling algorithm that approximates optimal inference.