Combined Blackbox and AlgebRaic Architecture (CBRA)

Combinatorial optimisation methods are concerned with an objective function defined over an (exponentially) large search space. Often these are solved with metaheuristics, such as simulated annealing, various evolutionary strategies, tabu search, and many others. However, such metaheuristics are generally rather static creatures with relatively simple behaviours and with most intelligence delegated to the heuristics, the neighbourhoods, selection/acceptance criteria and/or the local search methods that they oversee. This static simpleness has often paid off in terms of simplicity of implementation: Often the metaheuristic is merely tens to hundreds of lines of code, in comparison to maybe tens of thousands of lines to implement complex local search moves (with the associated data-structures needed for them to run quickly). However, the metaheuristics do have some decisions to make, and simple static decisions tend to lead to an over-reliance on careful tuning of parameters. There is a long-standing, but ever-growing, desire that metaheuristics become much more dynamic, self-adaptive, and with inbuilt learning mechanisms, in short, that they become more intelligent. There have been various attempts to add this intelligence. This short abstract cannot hope to do justice to them all, however, some pointers into the recent literature are: Hyper-heuristics (Burke et al 2003; Chakhlevitch and Cowling 2008; Burke et al 2009; Ross 2005; Özcan et al 2008), reactive search Battiti et al (2008), and self-adaptive genetic algorithms. This position paper is an attempt to give one view of how some different approaches are related. (Of course, this is just one view and is not intended to supplant other views.) The starting point is that optimisation methods can often be classified at a very high level according to whether they are blackbox or whitebox: