Up to now, only sequential compositions and single-level, template-based compositions based on hierarchically structured replacement patterns were considered for configuration. The evaluation of configurations and the learning of evaluations have also only been considered for these cases. In addition, the approaches to date assume that the configuration is completed before the services are executed.
These restrictions must be broken down for a distributed, dynamic market infrastructure. To this end, the following specific objectives are pursued:
- Sequential-hierarchical configuration
The two approaches to the composition of services, sequential and template-based, will be combined into a sequential-hierarchical approach. The core idea is to compose services in the form of sequences of appropriately instantiated templates. In this way, a good compromise between flexibility, expressiveness, reusability and complexity should be achieved.
- Interleaving of configuration and execution
The strict separation of the configuration phase from the execution phase is to be overcome. Instead, the configuration and evaluation of services should be interleaved with the execution. Such interleaving makes it possible, for example, to consider analyses of the data that have only been partially processed in the further composition.
As a concrete application domain for evaluation, AutoML is considered. AutoML deals with the automatic selection, parametrization, and aggregation of machine learning services. This field of application naturally requires an interleaving of configuration and design as performance on test sets can be used to guide search.