[2021-02] Learning from Failure: Exploration and Exploitation in Multistage Problem
[2021-02] Learning from Failure: Exploration and Exploitation in Multistage Problem
Jung, Kyungran
Moon Soul Graduate School of Future Strategy, KAIST
Kang, Heesuk
Moon Soul Graduate School of Future Strategy, KAIST
Kim, Jihyun
School of Business, Yonsei University
Yang, Jaesuk
Moon Soul Graduate School of Future Strategy, KAIST
Abstract
The canonical trade-off between exploration and exploitation has mainly studied the optimal choice strategy based on positive multi-peaks. Recently, many empiric studies have argued the importance and value of learning from failure. Therefore, we link the two fields of organizational learning and analyze how failure affects exploration and exploitation. In particular, we focus on the failures to be avoided due to catastrophic events or operational complexity and represent the complexity of failures in a multistage environment. We simulate the process of the agent solving the sequential decision-making problem with the Q-learning algorithm. We show that exploitation outperforms exploration in an environment with failure, and this tendency is prominent when a failed organization is attentive to survival. If an organization tries to learn from failure, we recommend local search and exploit internal knowledge resources in advance.
Keywords: learning from failure, exploration and exploitation, q-learning