• Using math.dot which does the dot product of two vectors using an internal 32-bit accumulator, then shifts right on the result before writing it back to a 16-bit integer.
• Using math.stat which computes the mean of an array using an internal 32-bit accumulator, then divides by the size of the array.
• Implementing a set of basic floating-point operations using native functions taking arrays of two 16-bit integers as parameters. That is not very clean but would work.
• Implementing a direct support for Lagrangian interpolation using native functions.

The first two suggestions are available on stock virtual machines, the last two requires adding new native functions.

I tried to use the range [-100, 100] for float, but the the final sum would be around 100000 and ASEBA can handle it.