Hrvoje Stojic, a researcher at the Department of Economics and Business, Universitat Pompeu Fabra, Barcelona will be visiting us this week and giving a presentation in the SWE Colloquium (abstract follows).

Exploration and generalization in multi-attribute decision making

Reinforcement learning (RL) models have been successful in explaining human and animal learning and decision-making. However, the experimental tasks often lack crucial characteristics of realistic decision situations: high-dimensional alternatives where features are informative of the alternative value. We developed a contextual multi-armed bandit task where participants choose repeatedly between multiple alternatives where payoffs were a function of two informative features. We report results of several experiments where the payoffs of the alternatives were governed by different type of functions. We find that although people seem to learn the function rather well, they do not exploit fully this knowledge and stick too much to the alternatives they know well. This is supported by the fit of the hybrid model that combines naïve RL model that ignores feature information and pure function-based RL model. Moreover, we examine exploration patterns of participants and find evidence that they are strategic in their exploratory choices: they choose alternatives to learn the function better, not necessarily to learn about the specific alternative.