Meta Planning Approach to Train Scheduling

Abstract

Computing the circulation plan for trains in a railroad is a hard problem. The number of variables considered, size of the search space and limitations on processing time define rigid limits to the solutions considered for any practical use. Our approach to train circulation planning [1] implements a simulation of discrete events that performs really well. Nevertheless, the approach is known to have some shortsightedness, specially when a global analysis is considered. To reduce the impact of such shortsightedness, a couple of solutions have been developed, including a second layer of intelligence on top of the core algorithm. This layer is called Adaptive Engine (AE), and is able to execute pre-processing and post-processing rules to fine tune the core engine’s output. The solution that is presented in this paper, called Meta Planning Engine (MPE), tackles another limitation of the tool, no matter how adjusted the implemented heuristics are, the core engine is computing only one output for any given input. With the MPE, multiple versions of the core engine execute in parallel and compute several outputs for any given input. We can even leverage the benefits of having different approaches running at the same time, without compromising the performance of the whole system. Preliminary results have shown it is possible to increase the solution quality by 7 to 10% while still maintaining the same range in terms of processing time.