RNNs reveal a new optimal stopping rule in sequential sampling for decision-making

Sequential sampling is a commonly studied decision-making environment where evidence accumulates over time until a decision boundary is reached. Normative theories of optimal stopping largely address homogeneous streams of evidence, where each time step carries the same amount of information. However, evidence from natural environments is often heterogeneous, with the informativeness of evidence varying over time; the optimal stopping rule under such conditions remains unknown. We trained recurrent neural networks (RNNs) to make decisions when receiving heterogeneous evidence streams and considering sampling costs and a time constraint inducing task urgency. In addition to replicating classic collapsing boundaries for stopping, we found a novel early commitment effect: the RNN adopts a lower decision boundary in the earliest time steps of decision-making. Normative analysis validated such a strategy as optimal. Examination of model policies showed that early commitment and collapsing boundary were driven by distinct mechanisms associated with sampling cost and time constraint, respectively. By bridging artificial networks and normative analysis, our work identifies early commitment as an optimal policy for decision-making in naturalistic environments.