Wang-Landau: adaptative random walk for fast convergence

Augustin Chevallier
Université de Nice Sophia Antipolis

I'll present here the Wang-Landau algorithm (to compute integrals), and an improved version. First I'll present the effect of the random walk on the convergence speed, then a random walk specifically designed to increase convergence speed. This random walk uses local geometric information to improve mixing time. Second, an important effect of high dimensionality on convergence speed will be discussed, along with a way to mitigate it.

Back to Complex High-Dimensional Energy Landscapes