No Free Lunch? (skip?)
NFL theorems - variants of old insight: most strings x of a certain size are incompressible.
That is, for most pairs of task descriptions (x,y): x does not convey any information about y. Assume uniform distribution on all tasks of a certain size: algorithm for solving random x usually no good for y.
But most real-world tasks are related! Uniform distributions are very special atypical cases.
Universal means: works optimally for all distributions.
OOPS exploits algorithmic task dependencies where possible, but does not perform miracles!
Back to J. Schmidhuber's OOPS page