Abstract:

At CQM we apply heuristics for various practical applications. For years, or decades even, the simulated annealing framework has been favored by us as a solution approach. Tailored simulated annealing more often than not outperforms other techniques that we try in early phases of our projects. However, a simulated annealing requires careful tailoring and implementation. For example, to ensure an elaborate search in the given limited time, one must have lightning fast evaluation on feasibility and potential value of a certain solution. Another challenge is choosing what and which neighborhoods to consider at various stages of the search. Expert knowledge and a lot of trial-and-error heavily guide all of these choices during development. This talk covers attempts on these challenges (and others) to integrate machine learning with simulated annealing to improve our optimization solutions. We also talk about their successes and challenges on practical applied client cases.