This trivially ensures convergence by allowing each weight to either only ever grow or only ever shrink during the optimization. This change should be backward compatible since it only changes behavior in cases that previously couldn't have converged anyway.
This trivially ensures convergence by allowing each weight to either only ever grow or only ever shrink during the optimization. This change should be backward compatible since it only changes behavior in cases that previously couldn't have converged anyway.
Fixes #20.