A search space 'cartography' for guiding graph coloring heuristics

摘要

We present a search space analysis and its application in improving local search algorithms for the graph coloring problem. Using a classical distance measure between colorings, we introduce the following clustering hypothesis: the high quality solutions are not randomly scattered in the search space, but rather grouped in clusters within spheres of specific diameter. We first provide intuitive evidence for this hypothesis by presenting a projection of a large set of local minima in the 3D space. An experimental confirmation is also presented: we introduce two algorithms that exploit the hypothesis by guiding an underlying Tabu Search (TS) process. The first algorithm (TS-Div) uses a learning process to guide the basic TS process toward as-yet-unvisited spheres. The second algorithm (TS-Int) makes deep investigations within a bounded region by organizing it as a tree-like structure of connected spheres. We experimentally demonstrate that if such a region contains a global optimum, TS-Int does not fail in eventually finding it. This pair of algorithms significantly outperforms the underlying basic TS algorithm; it can even improve some of the best-known solutions ever reported in the literature (e.g. for dsjd 1000.9).