Abstract
This paper is part of the study implementation of a new training algorithm for multi-dimensional wavelet networks called MDWNN-GA-MA using the genetic algorithm and multiresolution analysis to approximate and model 3D objects. This new approach aims at avoiding the weaknesses of old approaches such as the slowness and the difficulty in finding an exact reconstruction of objects especially when increasing the level of the decomposition. The result of the simulation reveals that this approach reduces the learning initialization cost and improves the gradient descent robustness. Indeed, multiresolution analysis has some interesting properties: such as starting with an object at high resolution, and generating several approximations can be generated. Details lost during the various stages of simplification can be returned if it requires greater precision. This technique speeds up the display surfaces and allows efficient compression.