Analysis and optimization of highly reliable systems

Abstract

In the field of network design, the survivability property enables the network to maintain a certain level of network connectivity and quality of service under failure conditions. In this thesis, survivability aspects of communication systems are studied. Aspects of reliability and vulnerability of network design are also addressed. The contributions are three-fold. First, a Hop Constrained node Survivable Network Design Problem (HCSNDP) with optional (Steiner) nodes is modelled. This kind of problems are N P-Hard. An exact integer linear model is built, focused on networks represented by graphs without rooted demands, considering costs in arcs and in Steiner nodes. In addition to the exact model, the calculation of lower and upper bounds to the optimal solution is included. Models were tested over several graphs and instances, in order to validate it in cases with known solution. An Approximation Algorithm is also developed in order to address a particular case of SNDP: the Two Node Survivable Star Problem (2NCSP) with optional nodes. This problem belongs to the class of N P-Hard computational problems too. Second, the research is focused on cascading failures and target/random attacks. The Graph Fragmentation Problem (GFP) is the result of a worst case analysis of a random attack. A fixed number of individuals for protection can be chosen, and a non-protected target node immediately destroys all reachable nodes. The goal is to minimize the expected number of destroyed nodes in the network. This problem belongs to the N P-Hard class. A mathematical programming formulation is introduced and exact resolution for small instances as well as lower and upper bounds to the optimal solution. In addition to exact methods, we address the GFP by several approaches: metaheuristics, approximation algorithms, polytime methods for specific instances and exact methods in exponential time. Finally, the concept of separability in stochastic binary systems is here introduced. Stochastic Binary Systems (SBS) represent a mathematical model of a multi-component on-off system subject to independent failures. The reliability evaluation of an SBS belongs to the N P-Hard class. Therefore, we fully characterize separable systems using Han-Banach separation theorem for convex sets. Using this new concept of separable systems and Markov inequality, reliability bounds are provided for arbitrary SBS.