| Current optimization techniques work well for single components represented by a single model. However, many of the problems we face today are multi-disciplinary problems requiring the integration of complex models from different fields to gain a more complete understanding of the overall performance of a biological, engineering, or human system. One example is a modern automobile. Multiple systems (such as the power train and electronic engine control system) are designed and built from various assemblies and components, all of which are then integrated into one final product. This design process evokes a systems-of-systems concept that is also found in agricultural facilities, aircraft design, and many other industrial applications where multiple systems are orchestrated to achieve common goals. Optimization of these complex systems is challenging. Tight coupling between the various models, discontinuous search spaces, and long run times can quickly defeat traditional optimization techniques.;Evolutionary algorithms provide a way to approach optimization of these complex systems. Evolutionary algorithms blend the information contained in a population of solutions to answer problems that thwart many classical optimization methods, but loss of diversity in the evolving solutions is a critical issue. As this information is shared between the population members, the diversity in that population decreases as the solutions converge to a single answer. For many challenging engineering problems this loss of diversity occurs too rapidly for novel solutions to emerge. In addition, systems of systems optimization problems are often deceptive because the global optimum is composed of multiple building blocks, making the preservation of diversity crucial.;This work presents graph based evolutionary algorithms as a tool to control the rate at which information is spread throughout an evolving population and thereby limit diversity loss. Graph based evolutionary algorithms impose a computational geography on the evolving population, placing barriers to information flow to allow for the development of the building blocks required to assemble one or more superior solutions. Graph based evolutionary algorithms can be used to find new solutions and decrease the time to convergence to a global optimum for complex, deceptive problems. In addition, the performance of a problem on a set of graphs can be used as a taxonomical character to classify evolutionary computation problems. If comparisons can be made between classified problems and a new problem being examined, it would be possible to select a graph that matches the desired performance. This careful graph selection can provide solutions that are both novel and superior to those found by standard evolutionary algorithms. Successful examples can be found in a variety of disciplines, including the engineering design problem of optimizing cook stoves for use in the third world to biological systems-of-systems, such as the tailoring of antibiotic regimens for use in swine production. |
