Relational Local Iterative Compression

Compression in the program space is of high importance in Artificial General Intelligence [Sol64, Hut07]. Since maximal data compression in the general sense is not possible to achieve [Sol64], it is necessary to use approximate algorithms, like AIXIt,l [Hut07]. This paper introduces a system that is able to compress data locally and iteratively, in a relational description language. The system thus belongs to the anytime algorithm family: the more time spent, the better it performs. The locality property is also well-suited for AGI agents to allow them to focus on ”interesting” parts of the data. The system presented here is to be opposed to blind generate and test approaches (e.g., [Sch04, Lev73]). On the contrary to the latter, it uses information gathered about the input data to guide compression. It can be described as a forward chaining expert system on relational descriptions of input data, while looking for the most compressed representation of the data. It is composed of a description/programming language, to describe facts (and a set of weights associated with each primitive of the language), local search operators, to infer new facts, and an algorithm to search for compressed global description. The relation operators and the search operators are domain-specific. Examples in the letter-string domain are given in the Experiments section. Due to lack of space, only a overview of the whole system can be given.