Bandits in the lab

Abstract

We experimentally implement a dynamic public‐good problem, where the public good in question is the dynamically evolving information about agents' common state of the world. Subjects' behavior is consistent with free‐riding because of strategic concerns. We also find that subjects adopt more complex behaviors than predicted by the welfare‐optimal equilibrium, such as noncut‐off behavior, lonely pioneers, and frequent switches of action.

Entry and exit decisions under public and private information: an experiment

“…This result suggests that there is a demand for information, and that participants do not exhibit a high degree of risk aversion since they are willing to explore the risky option. These results are also supported by other bandit experiments in the lab (Hoelzemann & Klein, 2021). 2 Our experimental design is motivated by the market entry decisions when information flows rapidly.…”

Exploration versus exploitation: A laboratory test of the single‐agent exponential bandit model

“…Our paper is related to the literature on bandit experiments where the risky action has a reward probability that can only take on one of two known values 2 . Hoelzemann and Klein (2018) analyze the game of strategic experimentation in Keller et al. (2005) in the lab and find that subjects appear to respond to strategic incentives by free‐riding.…”

“…Our experiment falls into the latter type of continuous time experiment. Most of the stochastic process experiments approximate either Brownian Motion (Anderson et al., 2010; Oprea, 2014; Oprea et al., 2009) or Poisson Processes (Hoelzemann & Klein, 2018; Hudja, 2019) in the lab. Our results suggest that subjects may deviate from theory in continuous time environments where subjects have to form beliefs of probabilities.…”

“…This simplification of the strategy space reduces the complexity of making decisions in this continuous time environment. Previous studies (Banks et al., 1997; Hoelzemann & Klein, 2018; Hudja, 2021) that have looked at experimentation environments have shown that subjects behave consistently with cutoff strategies in environments where cutoff strategies were not imposed.…”

“…This simplification of the strategy space reduces the complexity of making decisions in this continuous time environment. Previous studies(Banks et al, 1997;Hoelzemann & Klein, 2018;Hudja, 2021) that have looked at experimentation environments have shown that subjects behave consistently with cutoff strategies in environments where cutoff strategies were not imposed.10 This prevents subjects from accidentally switching to the safe action after the initial reward reveals that the subject has a good risky action.11 The longer sessions are the High Discount Factor Sessions, which have longer periods on average.12 The power analyses in Supporting Information S2: Appendix D suggest that our experiment is well-powered with this number of subjects.13 Smaller subsets of data can be used. For example, these results are robust to using the subset of data where the period lasts at least 250 ticks and where the state is bad or the first reward occurs after 249 ticks.…”