Seeking global minima (Q1322988)

The authors remark that neural networks can be used to solve problems in combinatorial optimization. In fact the dynamics of a neural network with symmetric connections result in the decrement of a configurational energy function. The paper presents a survey of some classical algorithms by comparing their performances for a specific problem. The binary representation problem was chosen where the configuration of the network is interpreted as a binary number, and a number is stored by having it to correspond to the configuration at minimum energy, thus allowing the global minimum to be known. Forteen algorithms are presented and compared in this interesting paper (examples: taboo research, network crystallization, simulated tunneling, simulated annealing, genetic algorithm, steepest descent,...). The conclusion is that the best rate of success is obtained with a field-ordered updating scheme. It gives the best solution with a minimal amount of neural processing steps.