Can I use parameters of different network training functions with each other?

If I am using trainscg, does it work to use parameters that are usually used for other network training functions? Like, can I use: net.performParam.regularization = B; net.trainParam.lr = A; For trainscg not just trainlm? Or does it have to be only these: (from https://www.mathworks.com/help/nnet/ref/trainscg.html)

• net.trainParam.epochs
• net.trainParam.show
• net.trainParam.showCommandLine
• net.trainParam.showWindow
• net.trainParam.goal
• net.trainParam.time
• net.trainParam.max_fail
• net.trainParam.sigma
• net.trainParam.lambda

I know it technically runs, but does it affect the result to use ones that are not in the above list/link?

Answer 1

[Greg Heath]

You can answer your own question via trial and error.

Greg

[Nis Jar]

I had been using trial and error and wanted to know if my improved results were random or not. So helpful.

[Greg Heath]

In order to make design comparisons valid, you should always initialize the RNG to the same initial state. Examples:

 rng(0)
rng('default') % Same as above
rng(4151941)   % A birth day

Greg

Answer 2

[Greg Heath]

My approach for successfully designing hundreds of nets is straightforward:

Accept all defaults except

 1. Initial state of the random number generator.

    rng(initial_state)  % Typically, I use rng(0)

 2. Outer loop over No of hidden nodes H:

    for h = Hmin:dH:Hmax

 % Typically, 10 equally spaced H values subject to the constraint 

    No. unknown weights+biases <= No. training eqns 

 3. Inner loop over Ntrials random initial weight settings

      for i = 1:Ntrials

 % Typically, Ntrials = 10. Rarely had to use 20 or 30

I have posted hundreds of examples in the NEWSGROUP and ANSWERS.

Thank you for formally accepting my answer

Greg