Optimal employee retention when inferring unknown learning curves

This paper formulates an employer's hiring and retention decisions as an infinite-armed bandit problem and characterizes the structure of optimal hiring and retention policies. We develop approximations that allow us to explicitly calculate these policies and to evaluate their benefit. The solution involves a balance of two types of learning: the learning that reflects the improvement in performance of employees as they gain experience, and the Bayesian learning of employers as they infer properties of employees' abilities to inform the decision of whether to retain or replace employees. Numerical experiments with Monte Carlo simulation suggest that the gains to active screening and monitoring of employees can be substantial.