Fig 1. Experimental design.

(A) Reward learning game. On each trial, subjects were asked to choose one of two circular images. Following their choice, subjects either received or did not receive a reward of 1 coin based on a fixed reward probability associated with the chosen image (0, 1/3, 2/3, or 1). Each game consisted of 48 such ‘learning’ trials, interleaved with 24 ‘testing’ trials wherein subjects chose between images about which they learned in prior sessions. Outcomes were not revealed in testing trials to prevent further learning. Every image first appeared in 64 learning trials over two sessions before subjects were tested on it. ITI: inter-trial-interval. (B) Daily schedule. Each day, subjects performed two experimental sessions on a specially designed mobile phone app [32] one in the morning (on average, at 8:56 am, and no earlier than 6:00 am) and one in the evening (on average, at 6:12 pm, and no earlier than 4:00 pm). In each session, subjects played two games in which they learned about a total of six images. (C) Experimental conditions. Four experimental conditions were implemented along 10 days of learning, each lasting two to three days. Each condition is illustrated via a representative selection of six pairs of images subjects chose between in two different sessions. Within the low concurrent diversity condition (left two columns), three images were learned over the span of each game, whereas in the high concurrent diversity (right two columns) six images were learned over the span of two games. Thus, in both conditions, each image appeared in 32 learning trials per session. In the low-cumulative-diversity conditions, each image was pitted against the same two images in two consecutive learning sessions, whereas in the high-cumulative-diversity conditions, images were pitted against different images (sometimes in two nonconsecutive sessions; see Methods for details). Three days of training, as opposed to two, were required for high-cumulative-diversity conditions so that each image could be pitted against different images in its two learning sessions. Conditions were randomly ordered, and equal in terms of average reward probability and number of images learned per session. Testing trials involved choosing between two images the subject already chose between during learning (Learned-pair trials, 25% of testing trials) or novel pairings of images learned separately (Novel-pair trials, 75% of testing trials). On average, pairs were tested 27 ±.8 times.