Statistical analysis of heuristics for evolving sorting networks

Abstract

Designing efficient sorting networks has been a challenging combinatorial optimization problem since the early 1960's. The application of evolutionary computing to this problem has yielded human-competitive results in recent years. We build on previous work by presenting a genetic algorithm whose parameters and heuristics are tuned on a small instance of the problem, and then scaled up to larger instances. Also presented are positive and negative results regarding the efficacy of several domain-specific heuristics.

Distributed island-model genetic algorithms using heterogeneous parameter settings

“…This is a classical problem in theoretical computer science [9] which has been used as a benchmark in evolutionary computation. We use the same genetic representation for sorting networks as Graham, Masum and Oppacher [10]. The 14-input sorting network problem is difficult and time consuming, because evaluating a single candidate individual (sorting network) requires executing the network on 2 14 test cases 1 .…”

“…In each run, the mutation rate and crossover rates were selected uniformly from [0.0, 1.0], and the population size was selected uniformly [10,100]. A standard binary chromosome representation was used, and roulette selection was used.…”

“…In our experiments, the mutation and crossover rates were unconstrained and selected uniformly from [0, 1.0], and the population parameter was constrained to be in [10,100]. Although we generally wanted the control parameters to be as unconstrained as possible, we chose an upper bound for 100 for the population parameter because each processor was running a generational GA, and it seemed that if we allowed the population size to be too large, the frequency of migration would end up being too small.…”