Machine Learning for Dynamic Incentive Problems
We propose a generic computational method for solving large-scale infinite-horizon, discrete-time dynamic incentive problems with hidden states. We first combine set-valued dynamic programming techniques with unsupervised machine learning to determine irregularly shaped feasible sets. Second, we generate training data from those pre-computed feasible sets to recursively solve the dynamic incentive problem by a supervised machine learning algorithm. Third, to speed up the time-to-solution process, we propose a generic parallelization scheme for dynamic incentive problems that allows an efficient use of contemporary high-performance computing hardware. This combination enables us to analyze models of a complexity that was previously considered to be intractable. To demonstrate the broad applicability of our computational framework, we study an insurance-like dynamic adverse selection problem with up to ten different, persistent types. Unlike the previous literature, we allow the agent to overreport his type. We find that the agent has to pay into the insurance for a shorter amount of time. This effect occurs because the agent now has access to the policies of a higher type than his own, which in turn forces the principal to prevent him from overreporting. Moreover, we observe that the overall time the agent has to pay into the insurance until he can file claims is reduced as the number of types increases.