Distributed hierarchical decision processes produce persistent differences in learning performance

Decision-making in social contexts often happens at multiple levels that operate on information with different degrees of abstraction, and at different time scales.  Such a way of organizing individual efforts is adopted by most of human organizations, independent from ownership, size, and purpose. Since Simon (1947), modern organization theory has emphasized the role of higher hierarchical levels in setting the decision premises that guide, constrain, and focus the operational tasks carried out by lower-level employees. Drawing on this perspective, we conduct a behavioral study and a hybrid model of multi-level learning, to examine the experimental behavior of hierarchical dyads, in which the high-level agent carries out strategy decisions, and the low-level agent operational ones.  Three main assumptions of our design make the task of the low- and high-level agents non-trivial: i) different stimulus features have different informative values for the low-level agent’s evaluation task; ii) the low-level agent evaluates each stimulus only based on the knowledge of a proper subset of stimulus features that the high-level agent makes available to her; iii) the informative value of the stimulus features is unknown to both agents: The high-level agent must learn to disclose the most informative stimulus features solely based on the feedback from the low-level agent’s performance. We show that such distributed hierarchical decision processes produce persistent differences in learning performance, which are neither based on differences in competences or match, but solely emerges because random initial conditions are reinforced.