ROL: line search in LSecantBAlgorithm

[ofmla]

What does the description in packages/rol/src/algorithm/TypeB/ROL_TypeB_LSecantBAlgorithm.hpp—"Provides an interface to run the line-search algorithm of Byrd, Lu, Nocedal, and Zhu (similar to L-BFGS-B)"—specifically mean? Does this implementation enforce both Wolfe conditions (sufficient decrease and curvature conditions), or does it relax the curvature condition, making it more similar to a backtracking strategy? Additionally, could you elaborate on the details of the line search implementation, including how step size selection is handled and whether it follows an interpolation-based approach like More-Thuente?

[cgcgcg]

@trilinos/rol

[dpkouri]

The current implementation employs a simple backtracking line search (along the lines of what Byrd et al. call their "first implementation" on page 1204). As Byrd et al. note, this can lead to the secant approximation not being updated. If there is a need, we can improve the line search using, e.g., More-Thuente.

[ofmla]

Many thanks for your reply. I'm a bit confused, and I apologize if this is a silly question. The algorithm implemented in the file is the l-BFGS method, so does the secant term in the file come specifically from what you mention about the line search process ?

Also, is there any ROL documentation available somewhere? I previously asked this question https://github.com/orgs/trilinos/discussions/13719 but didn’t receive a response.

[dpkouri]

Mechanically speaking, this is L-BFGS-B. However, we allow the user to employ any secant approximation, not just BFGS. ROL has default implementations of L-BFGS, L-DFP and L-SR1. The user can also provide their own secant implementation.