Be empowered: guiding adaptation through potential information flows

  • Daniel Polani (University of Hertfordshire, United Kingdom)
G3 10 (Lecture hall)


(joint work with Alexander Klyubin and Chrystopher Nehaniv)

The central importance of (Shannon) information as a resource for the adaptation of living organisms or agents has been increasingly established in the last years. The information perspective for the characterization of an agent's operation is highly attractive since it provides a universal currency and language for any kind of "information processing" taking place both within the agent and in the dynamics of its interaction with the environment, and allows one to adopt an extreme bottom-up view. Furthermore, it is "coordinate-free" in the sense that it allows to formulate principles and "balance sheets" without having to refer to a particular information processing mechanism

The use of Causal Bayesian Networks (CBNs) has been established as a successful technique to create informational models of agents and their perception-action loop. This technique allows the tracking of information flows through the composite agent-environment system, the generalization of Ashby's Law of Requisite Variety, or the application of generalized Infomax principles. In particular, the latter provide a path for the generation of structured information processing architectures with no assumptions beyond the agent being "embodied" in some structured environment. Phenomena such as active sensing emerge from the principle as a natural side effect.

The transparency of modelling the perception-action loop using the CBN formalism allows one to identify additional phenomena and quantities of interest. Specifically, in the present talk, I will introduce and discuss "empowerment" which is essentially the amount of potential information that an agent could inject into the environment via its actuators and recapture via its sensors. In the simplest of cases, this reduces to an agent-external channel capacity, but in general one requires CBNs to formulate empowerment.

In a situation where an agent has no prior preferences, its empowerment turns out to provide a utility which draws it to "interesting" states in the system. Since empowerment only depends on the embodiment of the agent, it can assign sensible preferences to states even in absence of any other prespecified drives (quantities of this kind we term "universal utility"). Understanding properties of possible universal utilities is particularly relevant for the success of adaptive systems, as the latter frequently have to be able to cope with novel situations that have not been previously encountered and for which the systems' innate drives are not appropriate or suitable drives may not exist yet at all.

I will show how, in a number of scenarios of varied quality and characteristics, the behaviour resulting from empowerment optimization is close to our intuitive expectations, sometimes achieved in a nontrivial way. In the discussion, I will suggest possible reasons for this and discuss lines for future research.

Antje Vandenberg

Max-Planck-Institut für Mathematik in den Naturwissenschaften Contact via Mail

Nihat Ay

Max Planck Institute for Mathematics in the Sciences, Leipzig

Ralf Der

Max Planck Institute for Mathematics in the Sciences, Leipzig

Mikhail Prokopenko

CSIRO, Sydney