Author(s)

J Blum & A Eskandarian

Abstract

Train-pathing for freight railroad traffic has defied traditional optimization methods. Intelligent agents, though, have shown promise in solving this problem for a portion of the railroad network. The agents include both domain-specific and generic problem solvers. Although domain-specific agents typically create a more robust search, certain generic methods, including genetic algorithms, increase the efficiency of many systems. This paper demonstrates how genetic agents can be made even more effective though the incorporation of domain-specific knowledge into their algorithms. The crossover operator in the genetic algorithm is designed to preserve meets and passes from parent solutions and to include the parent solution performance in the selection of the crossover point location. The agents were tested in a realistic simulation program on portions of the Burlington Northern and CSX railroads. The domain-specific genetic agents were successful as much as 33% more often than generic genetic agents. Furthermore, compared to runs without genetic agents, the domain-specific agents produced solutions that induced improvements that were on average 10% better than with generic genetic agents.

Keywords