Asymmetric learning and adaptability to changes in relational structure during transitive inference

Abstract

Humans and other animals can generalise from local to global relationships in a transitive manner. Recent research has shown that asymmetrically biased learning, where beliefs about only the winners (or losers) of local comparisons are updated, is well-suited for inferring relational structures from sparse feedback. However, less is known about how belief-updating biases intersect with humans capacity to adapt to changes in relational structure, where re-valuing an item may have downstream implications for inferential knowledge pertaining to unchanged items. We designed a transitive inference paradigm involving one of two possible changepoints for which an asymmetric (winner-or loser-biased) learning policy was more or less optimal. Participants (N=83) exhibited differential sensitivity to changes in relational structure: whereas participants readily learned that a hitherto low-ranking item increased its rank, moving a high-ranking item down the hierarchy impaired downstream inferential knowledge. Behaviour best captured by an adaptive reinforcement learning model which exhibited a predominantly winner-biased learning policy but also modulated its degree of asymmetry as a function of its choice preference strength. Our results indicate that asymmetric learning not only accounts for efficient inference of latent relational structures, but also for differences in the ease with which learners accommodate structural changes. Author SummaryWhen reasoning about relationships between objects, events, or people, humans can readily use previous experiences to infer relations that they have never encountered before. For example, if Anna beats Bruce at tennis, and Bruce beats Clara, then one can predict that Anna will likely also beat Clara. Human learning in such transitive inference problems tends to be winner-biased - that is, upon observing Annas victory over Bruce, a spectator would be more likely to attribute this outcome to Annas skill than to Bruces lack thereof. However, in a constantly changing world whose comparative relations are rarely static, humans must also be able to infer how changes in the outcomes of certain comparisons bear on other relationships within a transitive hierarchy. Combining behavioural testing and computational modelling, we show that a learning strategy that preferentially focuses on the winners of comparisons induces greater flexibility for certain types of hierarchy changes than for others. In addition, we provide evidence that humans may dynamically adjust their degree of learning asymmetry according to the current strength of their beliefs about the relations under comparison.