Self-assembly in Patterns with Minimal Surprise: Engineered Self-organization and Adaptation to the Environment

For complex and open-ended robot behaviors, it may prove to be important to find an intrinsic driver for pattern formation and self-organization. We apply methods of evolutionary computation and the idea of evolving prediction networks as world models in pair with action-selection networks to implement such a driver, especially in collective robot systems. Giving fitness for good predictions when evolving causes a bias towards easy-to-predict environments and behaviors in the form of emergent patterns, that is, minimal surprise. However, stimulating the emergence of complex behaviors requires to carefully configure allowed actions, sensor models, and the environment. While having shown the emergence of aggregation, dispersion, and flocking before, we increase the scenario's complexity by studying self-assembly and manage its feasibility by limiting ourselves to a simulated grid world. We observe emergent patterns of self-assembled robots adapted to different environments. Finally, we investigate how minimal surprise can be augmented to engineer self-organization of desired patterns.