Between heuristics and optimality: Flexible integration of cost and evidence during information sampling

Effective decision making in an uncertain world requires balancing the benefits of acquiring relevant information with the costs of delaying choice. Optimal strategies for information sampling can be accurate but computationally expensive, whereas heuristic strategies are often computationally simple but rigid. To characterize the computations that underlie information sampling, we examined choice processes in human participants who sampled sequences of images (e.g. indoor and outdoor scenes) and attempted to infer the majority category (e.g. indoor or outdoor) under two reward conditions. We examined how behavior maps onto potential information sampling strategies. We found that choices were best described by a flexible function that lay between optimality and heuristics; integrating the magnitude of evidence favoring each category and the number of samples collected thus far. Integration of these criteria resulted in a trade-off between evidence and samples collected, in which the strength of evidence needed to stop sampling decreased linearly as the number of samples accumulated over the course of a trial. This non-optimal trade-off best accounted for choice behavior even under high reward contexts. Our results demonstrate that unlike the optimal strategy, humans are performing simple accumulations instead of computing expected values, and that unlike a simple heuristic strategy, humans are dynamically integrating multiple sources of information in lieu of using only one source. This evidence-by-costs tradeoff illustrates a computationally efficient strategy that balances competing motivations for accuracy and cost minimization.