The field was developed by Andrey Kolmogorov, Ray Solomonoff and Gregory Chaitin starting in the late 1960s. There are several variants of Kolmogorov complexity or algorithmic information. The most widely used one is based on self-delimiting programs and is due to Leonid Levin (1974).
To formalize the above definition of complexity, one has to specify exactly what types of programs are allowed. Fortunately, it doesn't really matter: one could take a particular notation for Turing machines, or LISP programs, or Pascal programs, or Java virtual machine bytecode. If we agree to measure the lengths of all objects consistently in bits, then the resulting notions of complexity will only differ by a constant factor: if I1(s) and I2(s) are the complexitites of the string s according to two different programming languages L1 and L2, then there are constants C and D (which only depend on the languages chosen, but not on s) such that
In the following, we will fix one definition and simply write I(s) for the complexity of the string s.
The first surprising result is that I(s) cannot be computed: there is no general algorithm which takes a string s as input and produces the number I(s) as output. The proof is a formalization of the amusing Berry paradox: "Let n be the smallest number that cannot be defined in fewer than twenty English words. Well, I just defined it in fewer than twenty English words."
It is however straightforward to compute upper bounds for I(s): simply compress the string s with some method, implement the corresponding decompressor in the chosen language, concatenate the decompressor to the compressed string, and measure the resulting string's length.
The next important result is about the randomness of strings. Most strings are complex in the sense that they cannot be significantly compressed: I(s) is not much smaller than |s|, the length of s in bits. The precise statement is as follows: there is a constant K (which depends only on the particular specification of "program" used in the definition of complexity) such that for every n, the probability that a random string s has complexity less than |s| - n is smaller than K 2-n. The proof is a counting argument: you count the programs and the strings, and compare. This theorem is the justification for Mike Goldman's challenge in the comp.compression FAQ:
Similar ideas are used to prove the properties of Chaitin's constant
The minimum message length principle of statistical and inductive inference and machine learning was independently developed by C.S. Wallace and D.M. Boulton in 1968. MML is Bayesian (it incorporates prior beliefs) and information-theoretic. It has the desirable properties of statistical invariance (the inference transforms with a re-parameterisation, such as from polar co-ordinates to Cartesian co-ordinates), statistical consistency (even for very hard problems, MML will converge to any underlying model) and efficiency (the MML model will converge to any true underlying model about as quickly as is possible). C.S. Wallace and D.L. Dowe showed a formal connection between MML and algorithmic information theory (or Kolmogorov complexity) in 1999.