CRC 901 - Configuration and Rating (Subproject B2)

This subproject deals with the configuration and rating of services. In more detail, it aims at investigating basic configuration techniques and improving the latter's effectiveness as well as the efficiency. In the first funding phase (until June 30, 2015), the focus was on the development of basic algorithms for automatic service composition, i.e., for the configuration of sequential service composition workflows based on formal descriptions. [Read more about the first phase...] In the second funding phase (until June 30, 2019), these algorithms have been enhanced to enable the configuration of more complex workflows such as loops and branching structures. However, the main focus of the second funding phase was on abstracting from the configuration of single workflows to an automated service composition on a service level based on prototypes. Furthermore, the configuration of service compositions on a formal level was interleaved with the execution of the actual service compositions to additionally include dynamic non-functional properties.

In contrast to the first funding phase, where only workflows of single operation were composed to service compositions, the configuration algorithm composes entire services. Furthermore. the OTF Provider has distinct prototypes for different types of services defining the workflows. In this way, the OTF Provider configures single workflows in an implicit way as the workflows are encoded in the prototypes. The advantage of prototypes is that a generic cross-domain approach would be infeasible in the face of the search space complexity. Abstracting from single workflows offers great potential for interesting applications in different domains.

Another milestone of the second funding phase was to interleave the configuration algorithms working on formal descriptions with the execution of service candidates and to feedback dynamically measured non-functional properties to the configuration algorithm. HASCO was developed as a framework for this type of configuration and it is based on an AI planning approach called HTN planning. In HASCO services are represented in terms of components that can have two types of labels: provided interfaces and required interfaces, and a set of parameters. While provided interfaces promise to provide some functionality, required interfaces indicate that a dependency exists on another service and therefore requires another service that can satisfy this required interface with one of its provided interfaces. Parameters can stem from different domains, e.g. categorical, numeric, etc., and can be restricted to certain subsets of these domains, e.g. a numeric interval. Starting from a provided interface HASCO searches for a well performing solution. To determine the performance, it makes use of a benchmark which then executes the service composition to measure its dynamic non-functional properties. More details can be found on HASCO's tool page [link].

In order to speed up the search for a good solution, we incorporated the use of machine learning methods into the composition process in the form of dyad ranking. Within the scope of the configuration of image pipelines, we managed to learn a ranking function which was able to rank services based on a given service description with respect to their probable performance.

As a case study, the problem domain of automated machine learning has been considered where the aim is to choose algorithms and their parameters tailored to the user's data. Transferred to the OTF Computing context, the challenge is to configure a service composition tailored to the customer's needs. More specifically, it means that the OTF Provider aims at the provision of a service composition of machine learning basic services that generalizes well on the data provided by the customer. The case study is an interesting scenario because the associated search space of the possible service compositions is of complex structure, extremely large (>10⁴⁰), the evaluation of dynamic non-functional properties can be done automatically and the measured properties are of particular importance to the customer. Furthermore, dynamich non-functional properties, e.g. classification accuracy, cannot be determined for new data unless the service composition candidates are executed. This case study has also been used as a running example in the Proof-of-Concept (PoC) [link].