Year: 2015
Type: conference-paper
Abstract: Contextual Multi-Armed Bandit (CMAB) tasks are a\nnovel framework to assess decision making in uncertain\nenvironments. In a CMAB task, participants are presented\nwith multiple options (arms) which are characterized by\na number of features (context) related to the reward as-\nsociated with the arms. By choosing arms repeatedly\nand observing the reward, participants can learn about\nthe relation between context and reward and improve\ntheir decision strategy. We present two studies on how\npeople behave in CMAB tasks. Within a stationary en-\nvironment, we ?nd that participants are best described\nby Thompson Sampling-based Gaussian Process mod-\nels. This decision rule incorporates probability match-\ning to the expected outcomes derived from a rational\nmodel of the task and it is especially well-adapted to\nnon-stationary environments. In a dynamic CMAB task\nwe again ?nd that participants are best described by\nprobability matching of Gaussian Process expectations.\nOur ?ndings imply that behavior previously referred to\nas \\irrational" can actually be seen as a well-adapted\nstrategy based on powerful inference algorithms.\nKeywords: Decision Making, Learning, Exploration-\nExploitation, Contextual Multi-Armed Bandits
Source: MPG.PuRe (Max Planck Society)
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W2396959124