Sergey:
I don't know if you noticed and revised these, but in the most recent version (3.16.0), we found the following:
In the manual: (1) The manual entry for "autogkintegrate" lists the information for "odesolversolve" (2) Mention of MinCGSetGradientCheck() still exists in both the manual and comments in the code, but the routine has been removed from 3.16.0. (3) Minor errors ... all easily found and corrected: "launcn" → "launch", "traling" → "trailing", "trainig" → "training", "i.e ." → "i.e.", "i.e" → "i.e.", "inself" → "itself", "discripancy" → "discrepancy", "strucure" → "structure", "⋯intial⋯" → "⋯initial⋯", "it's" → "its" (this one requires careful analysis of the text to determine which ones are actually errors).
In test_c: (4) A long-standing hidden bug: in testablasunit_testsyrk(), cct is initialized with itself, instead of with the transpose of cc. (5) The routine BasicMultiLayerRBFTest1D (basicmultilayerrbf1dtest in the C++ version), which you disabled, generally works for errtype == 1, if setting max|y0 - y1|/b1 < 0.45 instead of to 0.35. Ratios at or beyond 0.45 have never been seen. For errtype == 2, the thresholds oralerr (0.10) and iralerr (0.01) almost always suffice, though one instance has been found at n: 100, ny: 2 with oralerr (0.10): 0.031056 0.028675 iralerr (0.01): 0.010313 0.006574
|