http://web.archive.org/web/20110208011537/http://neural-networks-at-your-fingertips.com/
i havent read the math, just the insidescience summary, but basically, i bet the idea is that they model 'power' or 'control' mathematically in terms of entropy ("capture as many future histories as possible"), and then for the cognitive entity to take actions that maximize this quantity in the future.
http://www.insidescience.org/content/physicist-proposes-new-way-think-about-intelligence/987
http://www.alexwg.org/publications/PhysRevLett_110-168702.pdf