Stability analysis of higher-order neural networks for combinatorial optimization.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Operations Research

Background:

• Recurrent neural networks with higher-order connections, termed higher-order neural networks (HONNs), are applicable to combinatorial optimization.
• A prior mapping of the traveling salesman problem (TSP) onto HONNs of arbitrary order established a family of networks for TSP solutions.

Purpose of the Study:

• To investigate the trade-off between network complexity and solution quality in higher-order neural networks for the traveling salesman problem.
• To analyze the stability of valid TSP solutions within higher-order neural networks of arbitrary order.

Main Methods:

• The study applies established stability analysis techniques to higher-order neural networks (HONNs) for the traveling salesman problem (TSP).
• The complexity of the HONN is systematically varied by adjusting the network order.

Main Results:

• Solution quality for the traveling salesman problem (TSP) improves with increased network complexity, specifically by raising the order of the higher-order neural network (HONN).
• The simplest network in the higher-order network family, the Hopfield network, is predicted to yield the lowest quality solutions for the TSP.

Conclusions:

• Increasing the order of higher-order neural networks (HONNs) enhances their effectiveness in solving the traveling salesman problem (TSP).
• The Hopfield network serves as a baseline, demonstrating that more complex HONNs provide superior solutions for combinatorial optimization tasks like the TSP.