Fault tolerant and dynamic evolutionary optimization engines

Abstract

Mimicking natural evolution to solve hard optimization problems has played an important role in the artificial intelligence arena. Such techniques are broadly classified as Evolutionary Algorithms (EAs) and have been investigated for around four decades during which important contributions and advances have been made. One main evolutionary technique which has been widely investigated is the Genetic Algorithm (GA). GAs are stochastic search techniques that follow the Darwinian principle of evolution. Their application in the solution of hard optimization problems has been very successful. Indeed multi-dimensional problems presenting difficult search spaces with characteristics such as multi-modality, epistasis, non regularity, deceptiveness, etc., have all been effectively tackled by GAs. In this research, a competitive form of GAs known as fine or cellular GAs (cGAs) are investigated, because of their suitability for System on Chip (SoC) implementation when tackling real-time problems. Cellular GAs have also attracted the attention of researchers due to their high performance, ease of implementation and massive parallelism. In addition, cGAs inherently possess a number of structural configuration parameters which make them capable of sustaining diversity during evolution and therefore of promoting an adequate balance between exploitative and explorative stages of the search. The fast technological development of Integrated Circuits (ICs) has allowed a considerable increase in compactness and therefore in density. As a result, it is nowadays possible to have millions of gates and transistor based circuits in very small silicon areas. Operational complexity has also significantly increased and consequently other setbacks have emerged, such as the presence of faults that commonly appear in the form of single or multiple bit flips. Tough environmental or time dependent operating conditions can trigger faults in registers and memory allocations due to induced radiation, electron migration and dielectric breakdown. These kinds of faults are known as Single Event Effects (SEEs). Research has shown that an effective way of dealing with SEEs consists of a combination of hardware and software mitigation techniques to overcome faulty scenarios. Permanent faults known as Single Hard Errors (SHEs) and temporary faults known as Single Event Upsets (SEUs) are common SEEs. This thesis aims to investigate the inherent abilities of cellular GAs to deal with SHEs and SEUs at algorithmic level. A hard real-time application is targeted: calculating the attitude parameters for navigation in vehicles using Global Positioning System (GPS) technology. Faulty critical data, which can cause a system’s functionality to fail, are evaluated. The proposed mitigation techniques show cGAs ability to deal with up to 40% stuck at zero and 30% stuck at one faults in chromosomes bits and fitness score cells. Due to the non-deterministic nature of GAs, dynamic on-the-fly algorithmic and parametric configuration has also attracted the attention of researchers. In this respect, the structural properties of cellular GAs provide a valuable attribute to influence their selection pressure. This helps to maintain an adequate exploitation-exploration tradeoff, either from a pure topological perspective or through genetic operations that also make use of structural characteristics in cGAs. These properties, unique to cGAs, are further investigated in this thesis through a set of middle to high difficulty benchmark problems. Experimental results show that the proposed dynamic techniques enhance the overall performance of cGAs in most benchmark problems. Finally, being structurally attached, the dimensionality of cellular GAs is another line of investigation. 1D and 2D structures have normally been used to test cGAs at algorithm and implementation levels. Although 3D-cGAs are an immediate extension, not enough attention has been paid to them, and so a comparative study on the dimensionality of cGAs is carried out. Having shorter radii, 3D-cGAs present a faster dissemination of solutions and have denser neighbourhoods. Empirical results reported in this thesis show that 3D-cGAs achieve better efficiency when solving multi-modal and epistatic problems. In future, the performance improvements of 3D-cGAs will merge with the latest benefits that 3D integration technology has demonstrated, such as reductions in routing length, in interconnection delays and in power consumption.