Unreliable heterogeneous workers in a pool-based evolutionary algorithm

Abstract

In this paper the effect of node unavailability in algorithms using EvoSpace, a pool-based evolutionary algorithm, is assessed. EvoSpace is a framework for developing evolutionary algorithms (EAs) using heterogeneous and unreliable resources. It is based on Linda’s tuple space coordination model. The core elements of EvoSpace are a central repository for the evolving population and remote clients, here called EvoWorkers, which pull random samples of the population to perform on them the basic evolutionary processes (selection, variation and survival), once the work is done, the modified sample is pushed back to the central population. To address the problem of unreliable EvoWorkers, EvoSpace uses a simple re-insertion algorithm using copies of samples stored in a global queue which also prevents the starvation of the population pool. Using a benchmark problem from the P-Peaks problem generator we have compared two approaches: (i) the re-insertion of previous individuals at the cost of keeping copies of each sample, and a common approach of other pool based EAs, (ii) inserting randomly generated individuals. We found that EvoSpace is fault tolerant to highly unreliable resources and also that the re-insertion algorithm is only needed when the population is near the point of starvation.