Unique Applications for Artificial Neural Networks. Phase 1

摘要

The investigation concerns the application of modular neural networks, workingsynergistically with genetic search, to provide a powerful means of intelligently controlling heuristic mathematical algorithms for large-scale vehicle routing and scheduling problems. The design lends itself naturally to parallel computing on loosely coupled networks of computers, and to implementation on parallel architectures such as MIMD machines. Extensive developmental work, coding and computational testing was carried on generic vehicle routing problems. The results are consistently superior to known alternatives, and provide strong motivation to extend the approach into more complex problem domains and military applications. The basic approach was also applied to routing problems with time constraints, a significant complication of considerable practical importance. Results of this problem are also consistently good, and there is potential to further investigate the use of the approach in this domain. Finally, very preliminary results are available for applying the methodology to routing and mission planning for remote autonomous military vehicles, such as Tomahawk cruise missiles or other smart weapons systems. In summary, the high performance achieved suggests that the multiparadigm approaches that utilize methods from artificial intelligence in conjunction with powerful and proven methods from mathematical combinatorial optimization can build upon the strengths of each constituent, and achieve performance that none of the methods can obtain in isolation.