SCCS Colloquium - Feb 26, 2020
|Date:||February 26, 2020|
|Time:||15:00 - 16:00|
Joachim Marin: Implementation of theFast-Multipole-Method Using AutoPas
AutoPas is a library aimed to provide efficient calculation of particle interactions in N-body problems for short-range forces. These forces converge quickly to zero for higher distances, such that only interactions between particles with short distances need to be calculated. This massively reduces the number of required calculations and therefore enables very fast computation of the forces between the particles. However, there are also long-range forces, which do not decay as fast, so that even interactions between particles that are far away from each other need to be evaluated. The Fast Multipole Method is an algorithm that can approximate these long-range forces accurately in linear time.
This thesis describes two approaches of how the Fast Multipole Method can be used with AutoPas and analyzes them in terms of performance and accuracy. The Fast Multipole Method will be implemented for the Coulomb potential, which is an example of such a long-range force.
Keywords: Fast Multipole Method, AutoPas
Rafael Hefele: Dimension Adaptive Efficient Global Optimization for Expensive Blackbox Problems
Bachelor's thesis submission talk. This is an external thesis at the TUM Department of Civil, Geo and Environmental Engineering,
advised by Koushyar Komeilizadeh and Prof. Fabian Duddeck and examined by Prof. Hans-Joachim Bungartz.
Growing demands on contradicting goals in the development of e. g. vehicle safety in passenger cars can be brought into accordance using design optimization for expensive blackbox problems. Safety features often require structural reinforcements while fuel efficiency is heavily influenced by weight and therefore calls for mass reduction. Efficient Global Optimization (EGO) is a prominent approach to tackle such problems. Its ability to reach improvement with a low number of expensive function evaluations, using a Gaussian Process (GP) based mathematical surrogate model, makes it especially useful when optimizing high-fidelity simulation models. EGO, however, also has its drawbacks. Although exploitation and exploration is balanced in the Expected Improvement (EI) infill criterion, EGO is not always able to lead to fast improvement or to find a near global minimum. In this work I introduce several strategies how to overcome these weaknesses by using prior knowledge on the influence of each design variable on the output function and exploiting their differences. Additionally, globally applicable measures are presented to further improve EGO’s performance which can be combined with the dimension adaptive options. Three main ideas are presented and discussed in this work:
- Use dimension adaptive parameters to build the surrogate model.
- Use a genetic sampling strategy on less important design variables to increase EI’s focus on more important dimensions.
- Introduce a second mathematical surrogate model with complementing properties to GP.
These different ideas are also intended to be used in combination. The approach is implemented in the package Dimension Adaptive Efficient Global Optimization (DAEGO) and applied to mathematical and physical test problems. The results show that the presented strategies can drastically improve EGO’s optimization performance in all cases without significant additional computational efforts.
Keywords: EGO, Expensive Blackbox Optimization, Expected Improvement