View source for Ray Solomonoff

'''Ray Solomonoff''' (1926–2009) was an American mathematician and the original discoverer of what is now called [[Algorithmic Information Theory|algorithmic information theory]] — a universal, substrate-independent measure of complexity based on the length of the shortest program that generates a given object. Solomonoff's formulation, developed in the 1960s, predated and encompassed the independent work of [[Andrey Kolmogorov]] and [[Gregory Chaitin]], though his early papers were less widely circulated.

Solomonoff's deepest contribution was '''universal induction''': a formal theory of prediction that combines [[Bayesian inference|Bayesian inference]] with algorithmic complexity to define an optimal predictor. The universal prior assigns higher probability to simpler hypotheses — those with shorter generating programs — and updates by conditionalization as data arrives. In the limit, this predictor converges to the true distribution faster than any other computable predictor, a result known as the '''optimality of universal induction'''.

This framework makes Solomonoff the foundational figure in theoretical machine learning — his universal induction is the mathematical ideal that practical learning algorithms approximate. The connection between compression and prediction, now central to fields from neural network training to data compression, was first made precise in Solomonoff's work. His later research explored the implications of algorithmic probability for artificial general intelligence, arguing that any sufficiently powerful learning system must approximate universal induction.

[[Category:Mathematics]]
[[Category:Computer Science]]