Fig. 1: Knapsack task and basic optimizing behaviors.
From: Computational complexity drives sustained deliberation

a, List of fractal images that symbolized the items in stimulus set 1 (left column), stimulus set 2 (middle column) and the items’ reward sizes (right column). b, Schematic representation of the knapsack task. The animals initiated the knapsack trials by touching a red central target. After initiation, an instance was displayed and remained on the screen for 5 s. During those 5 s, the number and identity of the selections were determined by the animal. When they selected an item, it was highlighted, it could not be deselected and the volume associated with the selection was added to the virtual knapsack at the top of the screen. If the sum of the items was less than or equal to 0.8 ml, then the sum was delivered at the end of the 5-s period. c, Schematic representation of a knapsack trial when animal exceeded 0.8 ml. In this case, no reward was delivered and a 4-s timeout was imposed. d, An additional fractal image was introduced as a ‘positive control’ that promised a reward equal to the knapsack limit. Bar graphs show the percentage of positive control trials when the animals selected the positive control item and nothing else. This response is optimal. Orange and brown bar graphs show data from monkeys G and B, respectively. The gray bar graph shows the percentage of positive control trials that an agent using a random sampling strategy chooses the positive control image and nothing else. Error bars are ±s.e.m. across n = 791 and 620 trials for monkeys G and B, respectively. e, An additional fractal image was introduced as a ‘negative control’ that promised no reward. Box plots show the percentage of negative control trials when the animals exhibited optimal behavior, here defined as not including the negative control item in the solutions. Orange and brown box plots show data from monkeys B and G, respectively. The gray dotted line shows the percentage of negative control trials that an agent using a random sampling strategy avoids including the negative control item. n = 36 and 26 sessions for monkeys G and B, respectively. Box plots show the median (line), quartiles (boxes), range (whiskers) and outliers (+).