Potential OOPS primitive: Genetic Programming
1985: Cramer invents GP: given some goal, use genetic algorithms / evolution to evolve tree-like or string-like programs. 1987: world’s second paper on GP, by undergrads Schmidhuber, Dickmanns, Winklhofer (TU Munich)
Most current GP implementations, however, ignore basic motivation for search in program space: no loops, no recursion! Therefore, in GP, 20 bit parity counts as a hard problem!
GP: no bias-optimality, no principled way of allocating runtime to programs & of reusing previous code.
First suggestion to make it bias-optimal: allocate runtime in proportion to mutation probability!
Unlike GP, OOPS is a self-improver: it also learns the programs that modify the programs (fixed in GP: crossover etc). But we may use GP as a primitive for OOPS, of course.
Back to J. Schmidhuber's OOPS page