Harmony Search-Based Approaches for Fine-Tuning Deep Belief Networks

Abstract

Harmony Search (HS) is a metaheuristic algorithm inspired by the musical composition process, precisely the composition of harmonies, i.e., the chain of different musical notes. The algorithm’s simplicity allows several points to improve to explore the entire search space efficiently. This work aims to compare different HS variants in image restoration using Deep Belief Networks (DBN). We compared standard HS against five variants: Improved Harmony Search (IHS), Self-adaptive Global Best Harmony Search (SGHS), Global-best Harmony Search (GHS), Novel Global Harmony Search (NGHS), and Global Harmony Search with Generalized Opposition-based learning (GOGHS). Experiments in public datasets for binary image reconstruction highlighted that HS and its variants obtained superior results than a random search used as a baseline. Also, it was found that the GHS variant is inferior to the others for some cases. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.