Asynchronous Master-Slave Parallelization of Differential Evolution for Multi-Objective Optimization

Abstract

In this paper, we present AMS-DEMO, an asynchronous master-slave implementation of DEMO, an evolutionary algorithm for multi-objective optimization. AMS-DEMO was designed for solving time-intensive problems efficiently on both homogeneous and heterogeneous parallel computer architectures. The algorithm is used as a test case for the asynchronous master-slave parallelization of multi-objective optimization that has not yet been thoroughly investigated. Selection lag is identified as the key property of the parallelization method, which explains how its behavior depends on the type of computer architecture and the number of processors. It is arrived at analytically and from the empirical results. AMS-DEMO is tested on a benchmark problem and a time-intensive industrial optimization problem, on homogeneous and heterogeneous parallel setups, providing performance results for the algorithm and an insight into the parallelization method. A comparison is also performed between AMS-DEMO and generational master-slave DEMO to demonstrate how the asynchronous parallelization method enhances the algorithm and what benefits it brings compared to the synchronous method.

Hybrid MPI-OpenMP parallel optimization for thermo-fluid problems

“…The obstacle setup is optimized by the means of evolutionary algorithm AMS-DEMO (Asynchronous Master-Slave Differential Evolution for Multi-objective Optimization) [10].…”

“…As shown in [10], the convergence ratio of the optimization procedure is inversely correlated to the number of parallel worker processes that are executing evaluations. In the presented case, the number of parallel worker processes equals the number of computers.…”

“…It is therefore expected that the convergence rate is lower for each added computer, yet measuring it exactly is much harder. In [10], it is shown that since the optimization is stochastic, the change of rate is only statistically significant when a large number of runs are made (more than 20), and when the number of parallel evaluations approaches the population size. For the presented case, since the population size is always larger than the number of computers and only a handful of repetitions are made for each test, the potential drop in the convergence rate is undetectable.…”

“…Figure 8 shows the average parallel speedup calculated from 3 repetitions of an experiment. Since the convergence rate is probably reduced but we are not sure by how much, we only show the weak-speedup [10], which only takes the number of performed evaluations into the account, but not the convergence ratio. The jagged appearance of the graph is due to the stochastic nature of the optimization, mixed with varying-length simulation, and the limited number of repetitions.…”

“…Optimizer implements the AMS-DEMO algorithm [10], which is an evolutionary algorithm [11] for multiobjective optimization [12] of real-valued functions. Evolutionary algorithms are a subgroup of stochastic optimization algorithms [13] and can solve problems, for which only the cost (or evaluation) function is known.…”