Integrating machine learning into heuristic optimization: How can we develop high-performing algorithms for real-world problems with graph-based representation?

Optimization problems are everywhere in our daily lives. Consider for example the truck deliveries of a wholesaler to its different national stores. Important decisions include the routes of each truck, the product types and quantities included in a delivery and the frequency of deliveries. These decisions are crucial for the wholesaler, since they have a direct impact on the product availability in its stores (and hence the sales), but also on the cost of the deliveries. To optimize such a complex problem, i.e. to find a delivery and loading schedule which ensures the highest profitability, algorithms have to be developed. In the retailer example, we furthermore need to consider tens of thousands of products, which implies that the problem is very difficult and requires a thoroughly thought out solution approach. As a result, any proposed algorithm should focus on finding a good heuristic solution, which may or may not be the best one, rather than focusing one's efforts solely on obtaining the best possible schedule. This in turn allows for quick decision making and the possibility to analyze the impact of different scenarios and assumptions. In the proposed research, I aim to develop heuristic algorithms, which incorporate knowledge gained from analyzing the large amount of data available.