Thomas Jatschka, Tobias Rodemann, Guenther Raidl, "Comparing Machine Learning Models in a Cooperative Optimization Approach for Distributing Service Points", EuroCAST, 2019.

We consider a variant of the facility location problem [2]. The task is to find an optimal subset of locations within a certain geographical area for erecting service points in order to satisfy customer demands as best as possible. This general scenario has a wide range of real-world applications. More specifically, we have the setup of stations for mobility purposes in mind, such as erecting bike sharing stations for a public bike sharing system, rental stations for car sharing, or charging stations for electric vehicles. When planning such systems, estimating under which conditions which customer demand can be fulfilled is fundamental in order to design and evaluate possible solutions [3]. To this end, demographic data is usually interlinked with geographic information, data on public transport, the street network, knowledge on manifold special locations, etc. Additionally, surveys of potential customers are performed. Customer demand information determined in such ways typically is vague, and not uncommonly a system built on such assumptions is not as effective as originally hoped for due to major deviations in reality. To possibly improve this situation, we investigate a cooperative optimization approach. More generally, interactive optimization approaches incorporate potential users on a large scale and more tightly into the data acquisition as well as the optimization process, for a recent review see [6]. A main problem with pure upfront surveys of potential users in scenarios like ours is that in general only incomplete information can be acquired this way. For example, users may specify where they would wish to have service points located in an ideal solution. However, there may be an arbitrary number of alternatives that also fulfill a user’s need to a certain degree. A promising optimization approach needs to be able to evaluate any potential candidate solution in a serious way. Instead of only acquiring demand information from potential users upfront, we therefore confront the potential users during the optimization with certain location scenarios and ask them how these would suit his needs and how possibly these scenarios can be improved to fulfill more of his demand. This feedback is used to incrementally gain more knowledge about how much demand may be fulfilled under which conditions. New, more promising candidate location scenarios can then be derived and again be presented to the users. The process is iterated on a large scale with many potential users and several rounds until a satisfactory solution is reached.