From T-mazes to labyrinths: Learning from model-based feedback

Abstract

Many organizational actions need not have any immediate or direct payoff consequence but set the stage for subsequent actions that bring the organization toward some actual payoff. Learning in such settings poses the challenge of credit assignment (Minsky 1961), that is, how to assign credit for the overall outcome of a sequence of actions to each of the antecedent actions. To explore the process of learning in such contexts, we create a formal model in which the actors develop a mental model of the value of stage-setting actions as a complex problem-solving task is repeated. Partial knowledge, either of particular states in the problem space or inefficient and circuitous routines through the space, is shown to be quite valuable. Because of the interdependence of intelligent action when a sequence of actions must be identified, however, organizational knowledge is relatively fragile. As a consequence, while turnover may stimulate search and have largely benign implications in less interdependent task settings, it is very destructive of the organization's near-term performance when the learning problem requires a complementarity among the actors' knowledge.