Issues with GaussianSI (unstable) & CV (extremely stable)

[tew42]

Hi @Zhiqiang-PANG ,

two more optimization algorithm issues/questions that I've come across...

1. The Gaussian peak percentage analysis appears quite unstable. When I look at optimization output for identical parameters, I get tables like the following. As you can see, all parameters are conserved across runs (as expected), EXCEPT GaussianSI.

exp	num_peaks	notLLOQP	num_C13	PPS	CV	RCS	GS	GaussianSI
1	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.548484848484848
2	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.536363636363636
3	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.527878787878788
4	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.545454545454545
5	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.541212121212121
6	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.549090909090909
7	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.527878787878788
8	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.521818181818182
9	3333	1455	472	7.04774211937663	0.00402016726116383	186.364053915599	939.25	0.532121212121212

2. On the other hand, over a rather large parameter space (and for two different analyses), the biggest variation in CV I have seen is less than 0.00025% . Is this supposed to be such a narrow interval? It doesn't seem very meaningful to normalize such a small range and give it somewhat similar impact as RCS or GS (where variation is easily > 100%).

[Zhiqiang-PANG]

Thanks for your comment. The CV is really dependent on the specific data. That's why there is less weight for CV in QcoE.

[tew42]

Any idea though why GaussianSI would be so unstable when running multiple times with same data & parameters?

[Zhiqiang-PANG]

Yes, this is caused by the random check on the peaks' gaussian fitting (does not check all peaks). This is designed to accelerate the optimization process without causing significant variation. I have find an approach to check all peaks but even faster than current strategy, is under testing and will be published, once stable in the coming months.