forum.alglib.net

I am running an extensive optimization (see the "BLEIC performance in v3.7.0 vs v3.6.0" thread for details).

With v3.6.0, my stopping criteria were: EpsG=1e-6, EpsX=0, EpsF=0, EpsO = EpsI = 1e-6, MaxIts=0 (as suggested in the BLEIC user guide and examples). At the expense of a longer calculation time, I could even make EpsG=1e-7 and the termination type in the result record was always 4 ("conditions on constraints are fulfilled with error less than or equal to EpsC").

With v3.7.0, I noticed that EpsG has to be significantly increased for the termination type to be 4 ("gradient norm is no more than EpsG"); otherwise, the termination type is 7 ("stopping conditions are too stringent, further improvement is impossible, X contains best point found so far").
In the worst-case scenario, I had to increase EpsG to 5e-6 (with N+1 linear constraints) or even 8e-6 (with N boundary and 1 linear constraints).
Therefore, there is no way to increase the accuracy of any solution in my set by lowering EpsG below these levels, even if more calculation time was allotted.

While I understand that the BLEIC algorithm has been improved in v3.7.0 to handle special cases, it looks like it requires much higher levels of EpsG and, consequently, may not yield as accurate a solution as in v3.6.0.

Comments?