Participants use a mixture of strategies to solve sequential decision problems

Full-information best choice problems typify many real-world decisions. Participants face a sequence of decision options and must accept or reject each option at the time it is encountered. The challenge is to use prior knowledge of the distribution of option values to predict whether continued sampling might result in a higher-valued option than the one currently on offer. On these tasks, biases are apparent when human decisions are compared against mathematical optimality solutions. Specifically, tasks involving choices about options presented as images appear to evoke an oversampling bias, in which participants perform suboptimally by rejecting optimal choices and continuing to sample new ones. Using computational modelling of ten datasets spanning five choice domains, we demonstrate two dominant sampling strategies. These strategies, which explained the overly long searches, involved a combination of overly optimistic predictions about future option values and an intrinsic perception that sampling itself has reward value. These strategies, and their associated systematic biases, could manifest in real-world decisions such as shopping, choosing dating partners or trustworthy job candidates.