Purchase this article with an account.
Timothy Vickery, Kyle Friedman, Rachel Bristol; Unreliable associations between visual features and values interfere with reward-based decision-making. Journal of Vision 2014;14(10):507. doi: https://doi.org/10.1167/14.10.507.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Many real world decisions depend on learning how visual characteristics of options predict their value. For instance, color is often predictive of food quality. Reinforcement learning (RL) models of such decisions often implicitly assume that only relevant, attended visual features play a role in decisions based on feature-value associations. How well are humans able to selectively attend to a given visual feature dimension to learn appropriate feature-value associations, while ignoring irrelevant feature-value associations? To assess this capacity, two experiments employed a 'four-armed bandit' task to assess whether learning of irrelevant feature-value associations intrudes on decision-making. In Experiment 1, four colors appeared on a screen in four locations, with the location of colors randomly determined on each trial. Participants selected a color option and then were monetarily rewarded or not based on a probability that varied independently for each option over time. Critically, reward probability was reliably associated with either color or location, and uncorrelated with the other dimension. Despite the fact that participants were fully informed of the relevant dimension, RL model comparisons suggest that they often inappropriately incorporated irrelevant and unreliable feature-value associations into their decisions. For instance, color-trackers inappropriately formed location-value associations, which played a role in decision-making. Since location was associated with motor responses, Experiment 2 randomized the binding of color, shape, and location of the 4 items, while participants attempted to track either color or shape while ignoring the other dimensions. Again, RL model comparisons suggested that participants often inappropriately incorporated irrelevant and unreliable feature-value associations into their decision-making (e.g., shape trackers inappropriately formed and employed color-value associations). These results suggest that visual feature-value associations are often formed even when the intent is to completely ignore a given dimension, perhaps as a means of maintaining flexible behavior in an uncertain environment.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only