Learning and Using Value Maps in Multi-Dimensional Choice

When making choices between multi-dimensional options, people draw upon knowledgeof the options’ various properties to make a decision. While much work has probed howindividuals make decisions in these modes, less is known of how learned value maps canbe flexibly composed to support novel decisions. Here we taught participants three separateperceptual-to-value mappings, each dividing a continuous perceptual dimension into five valuecategories. We then had them choose between composite options never encountered duringlearning while a mouse-contingent display forced serial information sampling. How the learnedmap was read during choice varied widely across people, from fully metric reads (e.g. howmuch better) to purely ordinal ones (e.g. which is ahead). We find that ordinal comparison isnot simply a degraded version of the same metric. Instead, participants mix between ordinaland metric reads. How well each dimension was learned set its weight, while how stronglylearning accuracy depended on the separation between categories lowered both the weightsand the metric share of the read. Both the readout and weights left separable signatures ininformation search. Learned value maps are thus composed and then read out to guide choice.