The Near-Term Liability of Exploitation: Exploration and Exploitation in Multi-Stage Problems

Abstract

The classic tradeoff between exploration and exploitation reflects the tension between gaining new information about alternatives to improve future returns and using the information currently available to improve present returns (March, 1991). By considering these issues in the context of a multi-stage, as opposed to a repeated, problem environment, we show that exploratory behavior has value quite apart from its role in revising beliefs. We show that even if current beliefs provide an unbiased characterization of the problem environment, maximizing with respect to these beliefs may lead to an inferior expected payoff relative to other mechanisms that make less aggressive use of the organization's beliefs. Search can lead to more robust actions in multi-stage decision problems than maximization, a benefit quite apart from its role in the updating of beliefs.

Keywords: Exploration and exploitation, maximization, multi-stage problems, reinforcement learning, softmax choice rule