Title : Discussing first implementation of Speculative Line Search (SLS)
with Youssef Diouane.
Presentation of my project. Getting insight on his research journey.
Idea 1 : Tune step length using 1D optimizatio argmin_{a>0} [ f(x+a.d) ] at each iteration.
Idea 2 : Idea 1 using dynamic model/surrogate.
Idea 3 : Dynamic scaling using Hessian approximation (inspired from CMA-ES algorithm). Using maximum eigenvalue as steplength for descent approx and its eigenvector as descent direction approx. Iterative construction of the Hessian ? (to be continued).
Discussing possible PhD opportunities. NSERC funding…
Readings: article on the scaling of Δ, δ | TREGO | MCS (smarter DiRect) | CMA-ES