Skip to content

Auto increase gamma #86

@aplavin

Description

@aplavin

Currently, the step size gamma parameter can only decrease, eg at

gamma, state.g_z = backtrack_stepsize!(

However, sometimes larger step sizes can be performed at some point during iteration.
I tried doing state.gamma *= 1.02 in the iteration loop, and seem to get ~2x faster convergence on some problems.
Gamma vs iterations plot looks like this:
image

Does this modification make sense in general? For some algorithms?
Should it be included into the package implementation itself?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions