Dietmar Heinke¹, Casimir Ludwig², Jordan Deakin^1,3; ¹University of Birmingham, UK, ²University of Bristol, Uk, ³Universität Hamburg

A dominant model of perceptual decision making assumes observers accumulate noisy evidence to a decision boundary. Theoretical work has shown that under some circumstances, the optimal strategy is to change the decision boundary dynamically while accumulating evidence. The majority of studies suggests that lowering the boundary (collapsing boundaries) over time would be optimal especially when evidence is difficult to obtain, or a decision deadline is involved. However, Malhotra et al. (2018) demonstrated that a certain mix of easy and difficult (but doable) decisions can actually lead to an increase of the decision boundary. This prediction was based on the following rationale: At the start of a trial, the decision boundary might be low and if the decision difficulty is low, it will be crossed rapidly. If the decision boundary is not crossed early, this itself is evidence that decision difficulty is high, and a higher decision boundary should be adopted. However, to our knowledge there is no evidence for such increasing decision boundaries. Here we present evidence from a decision task with RDK-stimuli of four noise levels (10%, 40%, 70%, 80%) and participants had to identify the direction of the coherent motion (left or right). We used Bayesian model fitting of a drift-diffusion model to identify whether and how the decision boundary varies with noise level. When the noise levels were blocked, decision boundary was larger for higher noise levels which might simply reflect “standard” adaptation of constant decision boundaries to decision difficulty. When the difficulties are randomly intermixed, such strategic adaptation is impossible. Nevertheless, we again found that the decision boundaries increased with the noise level. Hence observers may adjust their decision boundaries dynamically within a decision epoch, confirming Malhotra’s et al. prediction. Future studies need to test these findings further and fit formal models with raising decision boundaries.

Acknowledgements: The first author was supported by a grant from UK-ESRC ES/T002409/1.