The long tail in Bayesian optimal control in uncertain environments

Evolutionary algorithms and other biologically-inspired approaches to optimization rely on models that can be simulated in software. When calibrating a model to noisy and insufficient data, the single best-fitting model is often used. In contrast, Bayesian model averaging is known to give a better handle on uncertainty, but at the price of vastly more computation. This paper asks how much better, at how much more computation. The example problem uses the Bates model of stock price behaviour applied to barrier options, a problem similar to risk management with rainfall models.