8.04.2009

OOPS

The Optimal Ordered Problem Solver (OOPS - Schmidhuber, 2004) is also aimed at incremental learning. Every time a problem instance is solved, the successful program is stored on the input tape. For the following problem instance, two searches are run in parallel: half of the time is allocated to testing prolongations of the latest solver, the other half to fresh programs, starting from scratch. Both searches are conducted using backtracking to simulate a parallel execution of all candidate solvers. Time is allocated according to a distribution over programs, which is obtained multiplying the probabilities of the individual instructions. If the instruction set includes primitives that can modify these probabilities, then OOPS can in principle display meta-learning capabilities by finding programs that speed-up the search for future solutions.
Thanks, Remko
OOPS

Geen opmerkingen: