The availability of accurate location information of constituent nodes becomes essential in many applications of wireless sensor networks. In this context, we focus on anchor-based networks where the position of some few nodes are assumed to be fixed and known a priori, whereas the location of all other nodes is to be estimated based on noisy pairwise distance measurements. This localization task embodies a non-convex optimization problem which gets even more involved by the fact that the network may not be uniquely localizable, especially when its connectivity is not sufficiently high. To efficiently tackle this problem, we present a novel soft computing approach based on a hybridization of the Harmony Search (HS) algorithm with a local search procedure that iteratively alleviates the aforementioned non-uniqueness of sparse network deployments. Furthermore, the areas in which sensor nodes can be located are limited by means of connectivity-based geometrical constraints. Extensive simulation results show that the proposed approach outperforms previously published soft computing localization techniques in most of the simulated topologies. In particular, to assess the effectiveness of the technique, we compare its performance, in terms of Normalized Localization Error (NLE), to that of Simulated Annealing (SA)-based and Particle Swarm Optimization (PSO)-based techniques, as well as a naive implementation of a Genetic Algorithm (GA) incorporating the same local search procedure here proposed. Non-parametric hypothesis tests are also used so as to shed light on the statistical significance of the obtained results.