Accuracy of Neuro One learning algorithms

ACCURACY of NEURO ONE learning algorithms.






This table tests the efficiency of Neuro One regression algorithms. It uses data certified by the NIST and described on this page or on the NIST official web page.

Each line represents one of the examples provided by the NIST. They are organized in three groups: easy, medium and hard examples.

The Algorithm column indicates which regression algorithm has been the most efficient for the chosen example: Levenberg-Marquardt (LM), Quasi-Newton (BFGS) or the Gradient algorithm (Grad).

Start 1 and Start 2 represent the initialisation of the parameters as they were proposed by the NIST: Start 1 initialises the parameters far from the final solution, whereas Start 2 initialises them close to the final solution.

For each one of those two initialisations, the iteration column indicates the number of iteration done by Neuro One for the considered algorithm (Levenberg-Marquardt, BFGS, Gradient)

The number of significant figures indicate how accurate Neuro One is in retrieving the parameters certified by the NIST. It is the minimum number of figures in common between all the parameters calculated by Neuro One and those certified by the NIST.

To get more accurate information, click on the example’s name.


		Start 1: Hard Initialisation		Start 2: Easy initialisation
Easy Examples	Algorithm	Iterations	Nbr of significant figures	Iterations	Nbr of significant figures	Remarks
Misra1 a	L-M*	26	10	17	10
Chwirut2	L-M*	15	8	10	10
Medium Examples
Kirby2	L-M*	14	10	12	9
Hahn1	L-M*	31	9	17	10
MGH17	L-M*			36	8	1
Gauss3	L-M*	11	9	13	10
Misra1 c	L-M*	21	11	19	11
Misra1 d	L-M*	17	8	18	10
Hard Examples
MGH09	L-M	188	7	44	7
Thurber	L-M*	49	7	40	7
BoxBOD	Grad+L-M	10+22	8	20	9	2
Rat42	L-M*	15	10	14	10
MGH10	Grad+BFGS	20+2943	6	614	6	3
MGH10	L-M	6036	10	302	10
Eckerle4	L-M*	37	10	18	10	4
Rat43	L-M*	22	9	12	7
Bennet5	BFGS	1657	4	2672	4	5


Remark 1: No good result with "Start 1"

Remark 2: From "Start 1", 10 iterations have to be done with a gradient algorithm

Remark 3: From "Start 1", 20 iterations have to be done with a gradient algorithm

Remark 4: There are 2 possibilities for parameters beta1 and beta2. Depending on their initialisation, thay are either both positive or both negative

Remark 5: The number of iteration with L-M algorithm superior to 32000

*: Indicates the examples that have been made with Neuro One default settings

Read the complete result table.

Go to the NIST models and data sets.

Please contact us if you wish to know more about NEURO ONE learning algorithms.