BFGS uses a 'line-search' to find out the optimal step size it can take in the direction of the gradient. If the line search fails, it means that the algorithm cannot find a step size small enouogh such that the function decreases in the gradient direction. Are you sure your gradient information is accurate? Are you using a finite differences scheme to evaluate the gradient?
Date: Thu, 8 Jan 2015 15:49:27 -0800 (PST)
From: Ziyuan Lin <[hidden email]>
Subject: [SciPy-User] How does ?line search failed? affect the
To: [hidden email]
Message-ID: <[hidden email]>
Content-Type: text/plain; charset="utf-8"
I am using scipy.optimize.minimize for my optimization problem.
Specifically, I tried solvers "L-BFGS-B" and "TNC", but both give me
"Linear search failed"-like messages on my problem. What is the reason of
the failure of the line search on these solvers? Does it mean that the
final "optimal" value is potentially not optimal? Thank you.