Collective Learning Generally Overcomes Local Optima and Converges to the Global Optimum

Local minima represent a major problem for neural network learning procedures. In this article we present a new procedure, collective learning, that leads to improved global convergence. We have tested our procedure on several neural networks and on the multimodal functions proposed by De Jong and Rastrigin. In our tests we have reached a success ratio of 100 %. In addition we give a few remarks on the theorie of collective learning and give an estimate of the convergence behavior as well. Moreover, our procedure is very fast.