An Algorithmic Refinement of Maxent Induces a Thermodynamic-like Behaviour in the Reprogrammability of Generative Mechanisms

05/18/2018
by   Hector Zenil, et al.
0

Reprogramming a generative mechanism to produce a different object is associated with a cost. Here we use the notion of algorithmic randomness to quantify such a cost when reprogramming networks. We identify an asymmetry in a measure of reprogrammability, suggesting an analogy with a thermodynamic asymmetry. The principle of maximum entropy (Maxent) quantifies the evolution of entropy or the uncertainty during state transitions in systems confined to an equilibrium condition. Here we define a generalisation based on algorithmic randomness not restricted to equilibrium physics, based on both distance to algorithmic randomness and reprogrammability. We advance a constructive preferential attachment algorithm approximating a maximally algorithmic random network. Hence, as a refinement on classical Maxent, networks can be quantified with respect to their distance to a maximally algorithmic random network. Our analysis suggests that the reprogrammability asymmetry originates from its non-monotonic relationship to algorithmic randomness. Our analysis motivates further work on the degree of algorithmic asymmetries in systems depending on their reprogrammability capabilities.

READ FULL TEXT

Please sign up or login with your details

Forgot password? Click here to reset