Author Correction: Generalization guides human exploration in vast decision spaces

Charley M. Wu, Eric Schulz, Maarten Speekenbrink, Jonathan D. Nelson, Björn MederNature Human Behaviour 2020 · DOI

In a rare turn of events for scientific publishing, an author correction has actually strengthened the original findings of a study. The 2018 paper, "Generalization guides human exploration in vast decision spaces," published in Nature Human Behaviour, recently underwent a correction involving a coding error in simulations and a mathematical typo. Far from undermining the research, these fixes demonstrate an even tighter alignment between computational models and human behavior.

The Challenge of Vast Decision Spaces

How do humans make choices when faced with thousands of options? In "vast" decision spaces, it is impossible to try every possibility. The original research proposed that humans solve this exploration-exploitation dilemma by generalizing—learning the underlying structure of an environment to predict the value of unexplored options.

The authors hypothesized that humans use Gaussian Process (GP) regression with a Radial Basis Function (RBF) kernel to perform this generalization. This model assumes that points close to each other are likely to have similar rewards.

Correcting the Model

The correction addressed two primary issues:

  1. Mathematical Typo: The equation for the RBF kernel was written incorrectly in the text, though the correct version was used for the actual data analysis.
  2. Simulation Error: In Experiment 2, simulated models for "smooth" environments were accidentally tested using rewards from "rough" environments.

The corrected RBF kernel equation is: $k_{RBF}(\mathbf{x}, \mathbf{x'}) = \exp\left(-\frac{||\mathbf{x}-\mathbf{x'}||^2}{2\lambda^2}\right)$

Impact on Results

The most significant impact of these corrections is seen in the comparison between human performance and model simulations. In the original publication, there was a noticeable gap between the function-learning model and actual human data in smooth environments.

Original simulation results showing a gap between the function learning model (green) and human performance (pink) in smooth environments.

Upon fixing the simulation mismatch, the results showed a much closer fit. In the corrected figures, the function-learning model (green line) and human performance (pink line) are nearly overlapping, particularly in smooth environments where spatial correlation is high.

Corrected simulation results demonstrating a much tighter fit between the function learning model (green) and human performance (pink).

Conclusion

These corrections reinforce the idea that human decision-making is heavily guided by function learning. By accurately perceiving and exploiting spatial structures, humans can navigate complex environments far more effectively than simple mean-tracking models would suggest. The authors have updated their public GitHub repository to reflect these changes, upholding the principles of open science and reproducibility.

Summary updated

BibTeX
@inproceedings{wu2020author,
  author = {Charley M. Wu and Eric Schulz and Maarten Speekenbrink and Jonathan D. Nelson and Björn Meder},
  title = {{Author Correction: Generalization guides human exploration in vast decision spaces}},
  booktitle = {Nature Human Behaviour},
  year = {2020},
  doi = {10.1038/s41562-020-00958-w},
}
Figure