Morph II: A Universal Agent: Progress Report and Proposal


The Universal Agent differs markedly from other universal problem solvers such as Soar, in that it does not divide domains into "problems" and "methods" and require a "toolkit" of human supplied heuristics and algorithms. The Universal Agent is a more fundamental approach in which each domain is viewed as a variation of the Game of Abstract Mathematical Relations and relies on mathematical and statistical techniques for discovering and exploiting the inherent structure of these domains. As far as possible, human intervention has been limited to a simple mathematical description of a given domain.

The Universal Agent falls directly within the hierarchical reinforcement learning paradigm that has been gaining popularity in the last few years, but goes beyond: in this learning model, the system is responsible for abstracting its own features and patterns, developing its own learning modules, and employing transformations on the representation...