Evolutionary learning for relaxation labeling processes
Marcello Pelillo, Fabio Abbattista, Angelo Maffione · ARCA (Università Ca' Foscari Venezia) · 1993
Relaxation labeling processes are a class of parallel iterative procedures widely used in artificial intelligence and computer vision. Recently, a learning algorithm for relaxation labeling has been developed which involves minimizing a certain cost function with a gradient method. Despite the encouraging results obtained so far, the gradient algorithm suffers from some drawbacks that could prevent its application to practical problems. Essentially, these include the inability to escape from local minima and its computational complexity. In this paper we attempt to overcome the difficulties with the gradient procedure and propose the use of genetic algorithms for solving the learning problem of relaxation. Some results are reported which prove the effectiveness of the proposed approach.