Multi-attribute Decision-making is Best Characterized by an Attribute-Wise Reinforcement Learning Model

Choice alternatives often consist of multiple attributes that vary in how successfully they predict reward. Some standard theoretical models assert that decision makers evaluate choices either by weighting those attribute optimally in light of previous experience (so-called rational models), or adopting heuristics that use attributes suboptimally but in a manner that yields reasonable performance at minimal cost (e.g., the take-the-best heuristic). However, these models ignore both the possibility that decision makers might learn to associate reward with whole stimuli (a particular combination of attributes) rather than individual attributes and the common finding that decisions can be overly influenced by recent experiences and exhibit cue competition effects. Participants completed a two-alternative choice task where each stimulus consisted of three binary attributes that were predictive of reward, albeit with different degrees of reliability. Their choices revealed that, rather than using only the “best” attribute, they made use of all attributes but in manner that reflected the classic cue competition effect known as overshadowing. The time needed to make decisions increased as the number of relevant attributes increased, suggesting that reward was associated with attributes rather than whole stimuli. Fitting a family of computational models formed by crossing attribute use (optimal vs. only the best), representation (attribute vs. whole stimuli), and recency (biased or not), revealed that models that performed better when they made use of all information, represented attributes, and incorporated recency effects and cue competition. We also discuss the need to incorporate selective attention and hypothesis-testing like processes to account for results with multiple-attribute stimuli.