Optimize operator failure -- "Cannot reset network to a smaller learning rate"

keith_drakekeith_drake Member Posts: 11 Contributor II
I have a fairly simple process using the Optimize Grid operator and the Neural Network learner, It fails on the 5th iteration (256 total). I am varying two parameters, NN Momentum and NN Learning Rate.

Error message: "Cannot reset network to a smaller learning rate."

The next iteration would have been NN_Learning Rate = 0.267.

Any guidance? Thanks!

Keith


image

Answers

  • iesnaolaiesnaola Member Posts: 8 Contributor II

    Hi Keith,

    The same happened to me. As long as I keep the Learning Rate below 0.26 it works.

    I read in another thread (about 2010) that there was some kind of bug. I thought that by now this bug would be fixed but it seems it is not.

     

    BR

    Iker

  • keith_drakekeith_drake Member Posts: 11 Contributor II

    It just happened again. My process is very simple ... using NN to learn more about the Windowing operator. Seems the NN bug is alive!

  • MartinLiebigMartinLiebig Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,503 RM Data Scientist

    Hey,

    i checked the code, the error should only be thrown for negative learning rates? It's:

    if (learningRate <= Double.MIN_VALUE)
    throw new OperatorException("Cannot reset network to a smaller learning rate.");

    are you running 7.X? Could you maybe post your process?

     

    Best,

    Martin

    - Sr. Director Data Solutions, Altair RapidMiner -
    Dortmund, Germany
  • JEdwardJEdward RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 578 Unicorn

    I came across this problem when providing support.  I'm not sure if it is the same on your data, but try turning the parameter Decay on (if it is not yet set). 

    This should automatically decrease the learning rate to one where the error does not occur. 

     

    I also recommend changing the parameter on Optimise Grid to Ignore Errors (or put the NN inside a Handle Exception operator).  This will allow you to test all the parameters you want without worry whether a parameter setting might fail or not. 

     

     

     

     

Sign In or Register to comment.