We present a flexible and scalable computational framework integrating machine learning and optimization theory to solve dynamic adverse selection models with persistent private information and many types. Our approach reformulates the model into a numerically tractable structure that bypasses set-valued dynamic programming; we formally prove that, under verifiable conditions, this relaxation yields the solution to the original problem. The recast problem is solved via a parallelized value function iteration algorithm, where high-dimensional, nonlinear functions are approximated using Gaussian process regression combined with Bayesian active learning. We apply our framework to two previously intractable models: one with persistent hidden information involving up to ten types and another incorporating multiple persistent types and overreporting. Validation against known solutions and rigorous credibility measures confirms accuracy. Allowing overreporting significantly alters long-run contract outcomes, concentrating consumption away from extremes and smoothing utility promises over time.

This paper presents a new method for the analysis of moral hazard principal-agent problems. The new approach avoids the stringent assumptions on the dis-tribution of outcomes made by the classical first-order approach and instead only requires the agent’s expected utility to be a rational function of the action. This assumption allows for a reformulation of the agent’s utility maximization problem as an equivalent system of equations and inequalities. This reformulation in turn trans-forms the principal’s utility maximization problem into a nonlinear program. Under the additional assumptions that the principal’s expected utility is a polynomial and the agent’s expected utility is rational in the wage, the final nonlinear program can be solved to global optimality. The paper also shows that the polynomial optimiza-tion approach, unlike the classical approach, extends to principal-agent models with multi-dimensional action sets.