Evolving Diverse Collective Behaviors Independent of Swarm Density

摘要

There are multiple different ways of implementing artificial evolution of collective behaviors. Besides a classical offline evolution approach, there is, for example, the option of environment-driven distributed evolutionary adaptation in the form of an artificial ecology [2] and more generally there is the approach of embodied evolution [1,3,6]. Another recently reported approach is the application of novelty search to swarm robotics [5]. In the following, we report an extension of the approach of [7]. The underlying concept is an information-theoretic analogon to thermodynamic (Helmholtz) free energy [8]. The assumption is that the brain is permanently trying to predict future perceptions and that minimizing the prediction error is basically inherent to brains. This is defined by the 'free-energy principle' of [4]. The struggle for prediction success requires a complementary force that represents curiosity and exploration. In this abstract we present an extended method called diverse-prediction that rewards not only for correct predictions but also for each visited sensory state. This proves to be a better approach compared to the method prediction that was reported before[7].