MOSES

From AGIRI.org

Meta-Optimizing Semantic Evolutionary Search (MOSES) is an evolutionary algorithm developed by Moshe Looks based on ideas developed collaboratively with Ben Goertzel. It is described in some presentations available at

http://www.metacog.org

Roughly speaking, MOSES solves an optimization problem over program space by evolving a population of demes, each one of which in turn attempts to evolve problem solutions. There is a layer of deme evolution controlling the creation of new demes and the abandonment of unproductive ones. Each deme is in essence an iteratively growing program, which at any point in time consists of

a currently-fixed core, called the exemplar
a population of hypothetical extensions to the exemplar

The best of the hypothetical extensions to the exemplar are chosen to become part of the exemplar, and then the exemplar is iteratively expanded.

The way the space of hypothetical extensions is explored is roughly as follows:

first, a population of extensions is generated, and then normalized using an extension of Holman's Elegant Normal Form
then, probabilistic modeling (similar to what's used in BOA) is used to figure out what patterns characterize the fitter extensions
then, new hypothetical extensions are generated based on the probabilistic model, and added to the population, and Step 1 is revisited

MOSES was designed for use within the Novamente Cognition Engine, in tight integration with Probabilistic Logic Networks.

[edit]