Evolving prediction machines: collective behaviors based on minimal surprisal

Evolutionary algorithms can be applied to evolve controllers for single robots and similarly for groups of robots. Collective behaviors of groups of robots are investigated in the field of swarm robotics [1] which is the application of swarm intelligence to the field of robotics. An option is to to apply methods from evolutionary robotics to swarm robotics, that is, evolutionary swarm robotics [3]. In a standard approach a fitness function is used to reward behavioral features that are desired. In this paper we follow an alternative approach that generates collective behaviors without explicit selection for desired behaviors. We evolve agents, that mainly focus on predictions of their future perceptions, but still observe a number of different collective behaviors as a result. This approach is motivated by the hypothesis that perception is essentially a process of probabilistic inference—an idea that goes back to Helmholtz [4]. Following this concept, the main task of a brain is to figure out appropriate causes to its perceptions. Hence, the brain is interpreted as a ‘prediction machine’ that learns to model its perceptions. A mathematical framework by Friston [2] defines an informationtheoretic analogon to the thermodynamic (Helmholtz) free energy which is basically the prediction error here. Friston’s approach of a ‘free-energy principle’ might open up opportunities to formulate a unified brain theory [2].