Swarm Intelligence

The basic principle of the swarm intelligence algorithms is to divide complex calculations between multiple, simple executive agents. In these cases we use the emergent collective intelligence of groups of these simple agents (Bonabeau). These computer science algorithms are inspired by observations of real ant swarms, since they solve complex tasks by simple local behaviour and activities. The swarm is capable to solve complex problems, while at the same time each individual ant has no overall view of the situation.

With our swarm simulation implementations, various SI algorithms can be used to cluster data sets and display the virtual swarm environment with a two or three dimensional visualization. The applied algorithms are founded on the research of V. Ramos, J. Handl, M. Dorigo, E. Bonabeau, E. D. Lumer, B. Faieta and some further researchers of the swarm intelligence science community. Beyond this literature research in the SI domain we have implemented simulation systems inspired by the algorithms of the current research projects of Ramos, Dorigo et al. Furthermore we have developed various extensions to the original systems, such as dynamic pheromone evaporation, object switching while carrying another object, dynamic picking and dropping probabilities and varying border conditions as well as changes of the world dimensions in two and three dimensions as well as elegant solutions for some problems concerning orientation and boundary conditions. Some of these different extensions to the initial simulation systems are leading to measurable changes of the swarm behaviour, while other extensions have not changed the behaviour significantly. All extensions have been evaluated and are described in the chapters four, five and six of the project documentation (only available in German).

For the evaluation of the developed systems input data is essential. One could simply generate object-features that are randomly distributed over their particular dimensions. However this technique brings less benefit. What’s necessary here are some artificial generated object clusters. We therefore implemented a functionality for the flexible generation of gaussian distributed objects in n-dimensional feature-space. The configuration is done via self explaining config-files, where to specify the parameters of the distribution for the particular features of the objects. The algorithm can be used as a function included with the developed swarm systems as well as a stand-alone programm creating a text-file containing the object-data for further usage.

The results and the states of a swarm-clustering have to be measured and interpreted. So the need for robust entropy-measures arises. Therefore we have developed a set of measures namely sorting and connectivity, that can be applied in every step of the clustering-process. It provides us with information about the current state of the information infrastructure, that is the world of the swarm system. The gathered information is necessary for logging the performance in progression of the algorithms.

Here Sorting deals with the feature-space similarity of neighboring objects on the information infrastructure of the swarm. Connectivity measures the density distribution of the objects as well.

We have also successfully employed the introduced entropy-measures to achieve an increase in performance together with the algorithm proposed by Handl et al. 2003). Due to this, one can break the static temporal behavior, where the distinct algorithmic phases are switched at predefined iterations. We therefore define adaptive runtime-phases with its particular parameter-values and favored outcomes. The state of the system is permanently measured. If convergence against the desired state is reached, the algorithm proceeds to the next phase. For a detailed description of the simulations and results please refer to our projects documentation (only available in german).