Organic Computing: SRA contributions to OC research

Self-organization of large numbers of autonomous subsystems can lead to emergent behavior, i.e. the whole system might display macroscopic properties not visible in the subsystems. Although emergence can – by principle – not be exactly predicted we can nevertheless determine certain quantitative macroscopic properties of order of an emergent system. This allows for a reaction on system level to recognize and prevent negatively emergent behavior. We have shown that emergence can be quantified in terms of entropy differences.

Learning and self-organization have undesired side effects when used in safety-critical realtime systems: They need time, and they learn by trial and error, i.e. they might lead to unwanted, dangerous or illegal situations in the real world. We have developed a multi-level architecture based on the Observer/Controller pattern, which reacts in realtime on a low level by imposing behavior modifications with a limited “radius”. On a higher level, we use Evolutionary Algorithms for a behavior optimization, whose results are introduced into the real system only after a validation (sand-boxing).

In close cooperation with KIT Karlsruhe we have used our multi-level Observer/Controller architecture for a learning traffic light controller, which (1) adapts its switching patterns to the actual traffic requirements, and (2) cooperates with neighboring traffic light controllers to achieve a flexible progressive signal system (“green wave”) according to the current traffic situation. The architecture has also been used for the optimization of communication networks.