A Fourier AIXI Approximation (Universal AI)

Abstract

AIXI is a pareto optimal theory of artificial intelligence and uses Theory of Universal Induction along with control theory to achieve this task. The Universal Induction theory presents the concept of the Solomonoff's Universal prior for sequence prediction, though it is incomputable. The challenge of making a computationally efficient version of the prior (if not optimal) and thereby using it in AIXI is addressed in this paper. The proposed model is then simulated for environments manifesting as n-order Markov sources and a suitable extension to general systems has been suggested. The superiority of this time and length bounded prior, in terms of computational complexity, over methods using purely binary strings has been presented.