Nature inspired Algorithm-Based Fault Tolerance on Global Computing Platforms. Application to Symbolic Regression
Sébastien Varrette, Marek Ostaszewski, Pascal Bouvry · Open Repository and Bibliography (University of Luxembourg) · 2008
most powerful distributed computing systems in the world. Such an architecture bases on volunteer computing and on various forms of incentives, that makes it attractive to any cheater who wants to be rewarded with little or no contribution to the system. Cheating can be modeled as alteration of output values for some or all tasks of the program being executed, thus result-checking techniques are mandatory to cope with that kind of selfish behavior. Algorithm-Based Fault Tolerance (ABFT) [5] is an error detection technique where the scheme of the fault tolerance is tailored to the performed algorithm to make it resilient to a limited number of cheats. This article studies the ABFT aspect of distributed Evolutionary Algorithms (dEAs). EAs belong to a class of computation techniques based on the Darwinian theory of evolution that search simultanously for a whole group (called population) of solutions (called individuals). An interesting feature of EAs is that the population tolerates some of the bad quality solutions, which may in the long run guide the search process. EAs can be futhermore dis-tributed to improve the quality of the solutions and reduce the computation time.