When learning that Aretha had an older sister Erma and a younger sister Carolyn, we can easily infer that Erma was born before Carolyn. It is well established that humans can infer transitive orderings (e.g., Erma > Aretha > Carolyn) also for larger sets, only from the pairwise relations between its neighbouring elements. Adopting a simple reinforcement learning model, Ciranka, Linde-Domingo et al. show that inference of transitive orderings can be facilitated by a seemingly biased learning strategy, where observers only update their belief about one of the pair members (e.g., who is older) but not the other (who is younger). The experimental results indicate that humans use such asymmetric learning specifically in settings with sparse feedback (where transitive inferences are required), but not when full information about all pair comparisons is available. Asymmetric learning leads to a compressed representation of the orderly set (e.g., the age difference between Erma and Aretha may appear subjectively smaller than that between Aretha and Carolyn), which matches how humans typically perceive quantitative continua, such as stimulus magnitude or monetary value.
Asymmetric reinforcement learning facilitates human inference of transitive relations
CCBY Bank of Standardized Stimuli (BOSS)
Max Planck Institute for Human Development
