A Simulated Annealing Hyper-heuristic: Adaptive Heuristic Selection for Different Vehicle Routing Problems

One of the motivations of hyper-heuristic research is to investigate the development of adaptive decision support systems that can be applied to a range of different problems and different problem instances [5]. One possible approach is to dynamically adjust the preferences of a set of simple lowlevel heuristics (or neighbourhood operators) during the search. Hyper-heuristics have been used to address a variety of search problems, including educational timetabling [6, 2, 8, 7], scheduling [10], and packing and layout optimisation [1, 9]. In this research, we intend to further investigate an adaptation mechanism in the context of vehicle routing problems (VRP). VRP has numerous practical applications and it has been one of the most studied combinatorial optimisation problems in the past fifty years. Several variants of VRP have been addressed. Examples include capacitated VRP, VRP with time windows, multiple depot VRP, periodic VRP, stochastic VRP, and VRP with pickup and delivery. However, most existing algorithms for these problems are tailored for a particular problem variant and it is generally difficult to reuse the same algorithm for a different scenario without significant algorithm redevelopment or parameter-tuning. Recently, there are a few important research papers being carried out to improve the generality of algorithmic methods so that they can handle problems across more than one VRP variant [11, 12]. In both papers, the searches are based on combinations of “destruction-repair operators”, guided by some problem dependent penalty functions. In this abstract, we want to continue this research direction using a simulated annealing hyper-heuristic technique in order to investigate how the algorithm can intelligently choose between different neighbourhood operators (heuristics) according to the different problems. In particular, we consider two of the most popular variants of vehicle routing problems (capacitated VRP and VRP with time windows) and investigate the adaptation of the heuristic-selection mechanism across these variants and at different stages of the search. Specifically, we want to investigate: 1). How the hyper-heuristic adapts to these two types of vehicle routing problems by changing preferences of low-level heuristics and whether the hyper-heuristic can automatically identify heuristics that are particularly good for a given type of vehicle routing problem? 2). How the hyper-heuristic adapts its selection decision during different stages of the search when solving a particular problem instance?