Optimal Ordered Problem Solver

Schmidhuber, Juergen
Optimal Ordered Problem Solver
Machine Learning Journal 54, 211-254, 2004.; short version: NIPS 15, 1571-1578, 2003. (HTML - 200 KB)

Abstract: OOPS solves one task after another, through search for solution- computing programs. It is an incremental extension of Levin's non-incremental optimal universal search. OOPS bias-optimally exploits solutions to earlier tasks when possible, and learns to solve problems unsolvable by traditional reinforcement learners, such as Towers of Hanoi with many disks.