On Friday, 19th of June, Matthias De Cock, Wouter Minnebo and Sean Stijven will present results and insights of their reseach projects in the area of Computational Intelligence (promotor Katya Vladislavleva). Everybody is warmly welcome to attend this event in room G.005 (Campus Middelheim) from 11.00 a.m. till 12.30 p.m. The preliminary schedule and abstracts of the presentations are given below.
11.00-11.30: "Effective single-objective optimization with Swarm Intelligence" by Matthias De Cock
Particle Swarm Optimization is an optimization approach using a simple multi-agent system - a swarm of particles to optimize given optimization problem (with objective
function andconstraints given explicitly). A swarm of particles traverses the response surface of the objective function in a search for function's optima. This process can be seen as a simulation of the behavior of a flock of birds looking for food, moving individually but keeping the position and the state of the other birds in mind, allowing them to find their optimal habitat.
Particle Swarm Optimisation has proven to be a choice solution when attempting to optimise a wide variety of functions. It has been successfuly used for various optimization problems in the last decade, and yet stays a quickly evolving field due to growing complexity of the problems that computational science is daring to solve. In this bachelor thesis I provide an efficient implementation of the algorithm and proceed to apply various modifications attempting to accomplish an increase in speed / robustness and allowing the algorithm to identify local optima as well as the global optima.
11.30-12.00: "Multi-objective optimization with Swarm Intelligence" by Wouter Minnebo
Starting off with a short introduction about Particle Swarms for those unfamiliar with the field, this presentation is geared towards all audiences. With this background several selection schemes and boundary handling strategies will be discussed and compared in more detail, including proposed hybrids.
Since the Particle Swarm Optimization algorithm is stochastic by definition, an adequate result cannot be guaranteed for any single run. To improve reproducibility and quality of optimization results, we enhance the algorithm implementation by merging different runs together. We conclude with a demonstration of the developed software which embodies the presented material.
12.00-12.30: "GPU-based balancing of given multi-dimensional input-response data" by Sean Stijven
The analysis of multi-variate input-response data is crucial in applications of computational intelligence related to modeling measured data produced by complex systems. Because often the data is measured under difficult conditions, it is not well-balanced in a sense that it does not cover the input-output space in a balanced way. Regions of the data space that are oversampled can also slowdown the modeling process.
There exist a heuristic, the Simple Multi-dimensional Iterative Technique for Subsampling (SMITS), which can be used to balance such data sets. The problem with this algorithm is that it is slow for datasets of large size and high dimensionality. By executing the algorithm using the GPU (Graphics Processing Unit) we can gain significant speedups. The GPU has evolved into a highly parallel, many core processor with very high computational power and a high memory bandwidth. By using NVIDIA CUDATM(Compute Unified Device Architecture) it is possible create algorithms that execute on the GPU.
In this presentation we will first give an introduction to data balancing. Then we will show how CUDA works, how to create programs with it and what problems there can be when porting code to CUDA. We will continue by introducing the SMITS algorithm for GPU and give a short explanation of the used techniques. To conclude we will announce the achieved speedups and present the results.