To mitigate random initial weights in classifier design
1. Initialize the random number generator e.g., rng(0)
2. Design ~ 100 nets using a double loop over ~ 10 values of hidden nodes
j=0, h = Hmin:dH:Hmax, j=j+1 and ~ 10 random weight initializations
i = 1:Ntrials
3. For each design record
a. s(i,j) = rng % state of the rng
b. R2(i,j) and R2a(i,j)% Rsquared and degree-of-freedom adjusted Rsquared
c. Nerr(i,j) % Number of errors
4. Choose the net with the best combination of
a. Low validation error rate
b. Small number of hidden nodes
5. Details are in many of my posts. Search ANSWERS using (~17 hits)
patternnet Ntrials
Hope this helps.
Thank you for formally accepting my answer
Greg
