University of Houston

The objective of this PhD is to design efficient algorithms for low-cost computing systems to solve optimization problems of objective functions that are computationally intensive; that are stochastic with large noise ratio; that are defined on a search space of high dimensions; and that exhibit complex non-linear landscapes. This type of optimization problems is recurring in many scientific areas such as modeling biological or ecological systems. For instance, the Virtual Prairie (ViP) project presents multiple optimum design problems having the above properties. The ViP project aims at understanding the dynamics of clonal plant communities and designing function-specific prairies via modeling, simulation and optimization.

Efficiency is regarded as a compromise between execution speed and cost. Thus, on the one hand, the Volunteer Computing (VC) platform, which is virtually accessible to almost every scientist, was selected to satisfy the low-cost objective. On the other hand, Evolutionary Algorithms (EA) were adopted as the optimization tool most compatible with VC properties, thanks to their built-in embarrassing parallelism. Genetic algorithms, being among the most popular EAs, were deployed on top of the Berkeley Open Infrastructure for Network Computing (BOINC) chosen to be the middleware enabling the use of VC resources. Nevertheless, the execution speed of EAs can be affected by the volatility and unreliability of the VC compute nodes coupled with the synchronization required at each iteration of the algorithm. To controvert this phenomenon, customized scheduling techniques of BOINC tasks were designed, based on the concept of assigning the most important work to the fastest and most reliable resources. These methods were evaluated with Monte Carlo objective functions and achieved remarkable improvement of the algorithm performance.