Recommending PSO Variants Using Meta-Learning Framework for Global Optimization

Since inception, particle swarm optimization (PSO) has raised a great interest across various disciplines, thus producing a large number of PSO variants with respective strengths. However, a variant may perform variously on diverse problems, which leads to the risk of the algorithm selection of PSOs for a specific problem without prior knowledge. Hence, it is worth investigating a link between problem characteristics and algorithm performance. To address this issue, we propose a recommendation system of PSO variants for global optimization problem using meta-learning framework. Benchmark functions in the learning instance repository are pictured by meta-features to obtain characteristics and solved by the candidate PSO heuristics to gather performance rankings. k-NN method is employed to develop meta-learning system for recommending the predicted rankings of candidate PSO-variants. Results show that the predicted rankings highly correlate to the ideal rankings and achieve high precision on best algorithm recommendation. Besides, problem surface characteristics play a key role in recommendation performance, followed by sample point characteristics. To sum up, the proposed framework can significantly reduce the risk of algorithm selection.