Ensemble Methods for Optimisation
Workshop @ GECCO 2017, Berlin
Organisers: Emma Hart and Kevin Sim
In the field of machine-learning, ensemble-methods that combine decisions from multiple learning algorithms to make accurate predictions have been the focus of intense research for well over a decade, with a broad range of empirical results underpinned by a sound theoretical understanding. Ensemble methods have also found favour within the constraint satisfaction and satisfiability domains where they are commonly referred to as portfolio methods. In the latter case, portfolios tend to be composed from exact solvers, and are evaluated according to run-time metrics.
On the other hand, research in ensemble-methods using meta-heuristic algorithms – in which solution quality rather than run-time is the driving factor – lags behind machine-learning and satisfiability research in both theory and practice. Many fundamental questions remain with respect to how to construct, and design ensembles that will be addressed during the workshop:
- How should we select algorithms to form an ensemble?
- How large should the selection pool be – and where do we find algorithms to form the pool?
- Are automated algorithm generation techniques required to design new algorithms to provide a large enough pool?
- Machine-learning theory suggests that diversity between components is a key factor – what diversity measures can be used to successfully distinguish between meta-heuristic algorithms?
- How should the ensemble operate? Algorithms might collaborate, i.e. the computational budget is divided between algorithms within the ensemble, or cooperate, in that different algorithms solve different instances?
- What domains are ensemble methods best suited to?