A critical take on the role of random and local search-oriented components of modern computational intelligence-based optimization algorithms

Abstract

The concept of computational intelligence (CI)-based optimization algorithms emerged in the early 1960s as a more practical approach to the contemporary derivate-based approaches. This paved the way for many modern algorithms to arise with an unprecedented growth rate in recent years, each claiming to have a novel and present a profound breakthrough in the field. That said, many have raised concerns about the performance of these algorithms and even identified fundamental flaws that could potentially undermine the integrity of their results. On that note, the premise of this study was to replicate some of the more prevalent, fundamental components of these algorithms in an abstract format as a measure to observe their behavior in an isolated environment. Six pseudo algorithms were designed to create a spectrum of intelligence behavior ranging from absolute randomness to local search-oriented computational architecture. These were then used to solve a set of centered and non-centered benchmark suites to see if statistically different patterns would emerge. The obtained result clearly highlighted that the algorithm’s performance would suffer significantly as these benchmarks got more intricate. This is not just in terms of the number of dimensions in the search space but also the mathematical structure of the benchmark. The implication is that, in some cases, sheer processing resources can mask the algorithm’s lack of sufficient intelligence. But as importantly, this study attempted to identify some mechanics and concepts that could potentially cause or amplify this problem. For instance, the excessive use of greedy strategy, a prevalent measure embedded in many modern CI-based algorithms, has been identified as potentially one of these reasons. The result, however, highlights a more fundamental problem in the CI-based optimization field. That is, these algorithms are often treated as a black box. This perception cultivated the culture of not exploring the underlying structure of these algorithms as long as they were deemed capable of generating acceptable results, which permits similar biases to go undetected.