The guilty net for the travelling salesman problem

A new, adaptive neural structure is proposed for solving the traveling salesman problem. While non-adaptive (Hopfield) networks were the first neural networks to be used for this problem, their practical limitations and dependence on hard-to-determine parameters motivated investigation into dramatically different neural approaches. The new approaches exploit adaptive neural networks, and outperform Hopfield type approaches by a substantial amount, but usually require thousands of iterations and additional mechanisms to ensure generation of a valid tour. The guilty net, an even more recent adaptive approach, utilizes a straightforward competitive learning algorithm with fixed neighbors and ‘conscience’ to automatically generate valid, short tours in only a few hundred iterations.