These methods are based on a uniform pseudo-random sampling of the parameter space. If their principle is not new, they gain success amongst the geophysicists during the last 20 years, due to the increasing power of modern computers. The question addressed by such methods is not only finding the model with the best data fit but also to retrieve information about the resolution power of a particular application. This area of statistical inference is reviewed for example by Edwards (1992); Sambridge (1999b); Mosegaard and Tarantola (1995). The role of prior information is investigated by all these authors but especially by Scales and Tenorio (2001). The parameter space does not generally extend to 
like in the case of iterative methods but it is restricted to a volume defined by the parameter prior ranges. All generated models are always confined in this volume.
When the dimensionality of the parameter space increases, the basic random generation of models becomes totally inefficient. This is why several refined approaches were developed during the last two decades, for instance the simulated annealing (Sen and Stoffa (1991); Rothman (1985)) and the genetic algorithms (Yamanaka and Ishida (1996); Boschetti et al. (1996); Stoffa and Sen (1991); Lomax and Snieder (1995)). There are also many variants of these methods, combining them with neural networks or with gradient methods (e.g. Boschetti and Moresi (2001); Chunduru et al. (1996); Devilee (1999)). The objective of these techniques is to seek a model with a globally optimal data misfit value. These methods and their variants usually need empirical tuning of several parameters that control the inversion process, ensuring computational efficiency and robustness against entrapment in local minima.
Recently, Sambridge (1999a) proposed an entirely different method based on the partition of the parameter space into Voronoi cells2.6( neighbourhood algorithm). It has only two tuning parameters and it is claimed as self-adaptive in searching a parameter space. The objective, which is different from the previously mentioned methods, is to sample (in an optimal situation) all the regions of the parameter space where models with acceptable data fit are found. This last technique has been chosen for our dispersion curve inversion tool. Its principles are examined with more details in the next section.