Anat Harari

Anat Harari is a Senior Research Associate in the Games, Dynamics and Strategic Control Division at the Institute for Advanced Dynamic Uncertainty. She holds a PhD in Electrical and Computer Engineering from the Technion – Israel Institute of Technology, where her doctoral research developed convergence theory for policy gradient algorithms in continuous-state stochastic control problems. Her thesis established non-asymptotic convergence rates for a family of natural policy gradient methods applied to discounted Markov decision processes with function approximation, and provided the first rigorous sample-complexity bounds for model-free approximations of the Hamilton-Jacobi-Bellman equation in controlled diffusion settings.

Harari's research sits at the mathematically precise junction between reinforcement learning and stochastic optimal control. She is interested in the conditions under which approximate dynamic programming methods inherit the structural properties of the underlying control problem — monotonicity, convexity, value function regularity — and in quantifying the degradation of these properties under approximation error and distributional shift. Her methodological contributions include variance-reduced policy gradient estimators for continuous-time systems and a class of entropy-regularised value iteration schemes with provable convergence to the viscosity solution of the limiting HJB equation.

At IADU, her work focuses on learning algorithms for Nash equilibria in mean field games, the design of scalable multi-agent reinforcement learning methods for large-population differential games, and the theoretical analysis of fictitious play and best-response dynamics in the mean field limit.