How to choose RandomDistributedScalarEncoder settings

Hey guys,

If I have a time series that has a minimun value of 8 and a maximum value of 20 and a precision of 0.01, what are the best parameters to use with RandomDistributedScalarEncoder for n, w, min, max, radius, and resolution for prediction with a target of 1% of average error?

I’m trying to achive a good solution and that’s my reasonig:

Since my precision is 0.01, my resolution must be 0.01.

If my target average error is 1% then my absolute error should be in the rounds of 0.12. Then I’ve chosen this value to be my radius parameter.

For the n and w values, my reasonig is that my range is 20-8=12 and my precision is 0.01, then i’m going to have 1.200 possible values to encode. So I have to have enough bits in my SDR to encode all those values and keep it sparse. So I choose n=3000 and w=21.

Is that a reasonable aproach?

1 Like

The rule I used is this:

  1. Calculate the number of buckets you want your input to be separated into
  2. Set RDSE width to: max(0.001, (maxInput - minInput) / buckets)
    Where maxInput and minInput are the maximum and minimum possible values for your input.
1 Like

Thank you @Jonathan_Mackenzie. But how do you calculate the number of buckets? What are the criterias that you use?

It’s a bit of guess work or whatever makes sense for the problem domain.

Thank you