Abstract

The ability to estimate and control extreme tail risks, besides being an integral part of quantitative risk management, is central to running operations requiring high service levels and cyber-physical systems with high-reliability specifications.  Despite this significance, scalable algorithmic approaches have remained elusive: This is due to the rarity with which relevant risky samples get observed, and the critical role experts need to play in devising variance reduction techniques based on instance-specific large deviations. Our goal in this talk is to describe a variance reducing importance sampling technique which can be flexibly combined with stochastic approximation to yield significantly faster algorithms for minimizing tail risks. We aim to bring out  how the central challenge of selecting a good importance sampler, which in turn depends on the knowledge of the solution we are seeking, can be tackled with a novel black-box importance sampler.