Risk-Averse Optimization (CVaR across 3 Domains)
I applied CVaR-based risk-averse optimization to nonlinear programming, perturbed SVM classification, and wine-fermentation model predictive control.
Under severe temperature mismatch between model and reality, CVaR control cut constraint violations from 100% of scenarios to 0%
I implemented Conditional Value at Risk (CVaR) objectives, which optimize for the worst-case tail of outcomes, and tested them on three problems: a noisy benchmark function, an SVM trained on perturbed data, and a wine-fermentation controller.
In my open-loop fermentation comparison with severe temperature mismatch, the CVaR controller reduced the violation rate from 100% to 0%, while using more control effort.
In a separate closed-loop 21-day hot-scenario simulation, I found that CVaR improved final sugar and ethanol outcomes at higher control cost. This result is distinct from the open-loop violation comparison.