Toward self-optimization of machine intelligence
R Michael Perry · 1984
For a long time there has been interest in putting a computer program that solves problems to work on the of its own abilities, though little has been done to formally realize this ambition. This work offers a mathematical theory of computerized problem solving. Under the theory a problem solver, represented as a computer program, is able to rigorously consider the of its own ability to problems, or in some sense, its The notions of and of what it means to solve it had to be broadened considerably beyond their usual meanings. In the resulting theory of a binary preference relation or ordering is assigned to a class of objects. The problem then becomes one of constructing an optimizing of objects that tend to an optimal condition in that, under the preference relation, only a finite number of terms of the sequence can be equivalent or inferior to any given, suboptimal object. The theory is able to handle the more conventional versions of problem solving such as theorem proving as well as esoteric forms of optimization in the limit. It is possible to define an IQ test for a problem solver or that solves limiting optimization problems. This test contains a number of problems to work on and measures the average performance of the organism. In this way a preference relation based on one interpretation of can be established. An organism is then able to construct an sequence of organisms that tend to a condition of maximal intelligence. A computer implementation of an organism is offered and results of attempted self-optimization of intelligence are shown. This represents a small part of the infinitely many operations that would be necessary to fully carry out a limiting optimization. The improved organism is constructed by and from the original by mutating its code. It is found to have modest but robust gains in performance suggesting that an improvement in general abilities has really occurred.