Approximation results for probabilistic survivability

Abstract

As multiagent systems (MASs) are increasingly used in industrial applications, the need to make them more robust and resilient against disruption increases dramatically. Kraus et al. [1] has developed a probabilistic model (assuming complete ignorance of dependencies between node failures) of survivability based on deploying each agent in a MAS on one or more nodes. Finding a deployment that maximizes survivability is highly intractable for two reasons: firstly, computing the survivability of any deployment is intractable, and secondly, going through an exponential number of deployments to find the best one adds another layer of intractability. In this paper, we study what happens when node failures are independent. We show that computing survivability in this environment is still intractable. We propose various heuristics to compute the survivability of a given deployment. We have implemented and tested all these heuristics. We report on the advantages and disadvantages of different heuristics in different environmental settings. * Sarit Kraus is also affiliated with University of Maryland.