Automatic online algorithm selection for optimization in cyber-physical production systems
| dc.contributor.advisor | Rudolph, Günter | |
| dc.contributor.author | Fischbach, Andreas | |
| dc.contributor.referee | Bartz-Beielstein, Thomas | |
| dc.date.accepted | 2023-07-10 | |
| dc.date.accessioned | 2024-02-07T06:58:35Z | |
| dc.date.available | 2024-02-07T06:58:35Z | |
| dc.date.issued | 2023 | |
| dc.description.abstract | Shrinking product lifecycles, progressing market penetration of innovative product technologies, and increasing demand for product individualization lead to frequent adjustments of production processes and thus to an increasing demand for frequent optimization of production processes. Offline solutions are not always available, and even the optimization problem class itself may have changed in terms of the value landscape of the objective function: Parameters may have been added, the locations of optimal values and the values themselves may have changed. This thesis develops an automatic solution to the algorithm selection problem for continuous optimization. Furthermore, based on the evaluation of three different real-world use cases and a review of well-known architectures from the field of automation and cognitive science, a system architecture suitable for use in large data scenarios was developed. The developed architecture has been implemented and evaluated on two real-world problems: A Versatile Production System (VPS) and Injection Molding Optimization (IM). The developed solution for the VPS was able to automatically tune the feasible algorithms and select the most promising candidate, which significantly outperformed the competitors. This was evaluated by applying statistical tests based on the generated test instances using the process data and by performing benchmark experiments. This solution was extended to the area of multi-objective optimization for the IM use case by specifying an appropriate algorithm portfolio and selecting a suitable performance metric to automatically compare the algorithms. This allows the automatic optimization of three largely uncorrelated objectives: cycle time, average volume shrinkage, and maximum warpage of the parts to be produced. The extension to multi-objective handling for IM optimization showed a huge benefit in terms of manual implementation effort, as most of the work could be done by configuration. The implementation effort was reduced to selecting optimizers and hypervolume computation. | en |
| dc.identifier.uri | http://hdl.handle.net/2003/42307 | |
| dc.identifier.uri | http://dx.doi.org/10.17877/DE290R-24144 | |
| dc.language.iso | en | de |
| dc.subject | Parameter optimization | en |
| dc.subject | Automatic algorithm selection | en |
| dc.subject | Benchmarking | en |
| dc.subject | Surrogate models | en |
| dc.subject | Cyber-physical production systems | en |
| dc.subject | System architecture | en |
| dc.subject | Industrie 4.0 | de |
| dc.subject.ddc | 004 | |
| dc.subject.rswk | Industrie 4.0 | de |
| dc.subject.rswk | Produktionssystem | de |
| dc.subject.rswk | Parametrische Optimierung | de |
| dc.title | Automatic online algorithm selection for optimization in cyber-physical production systems | en |
| dc.type | Text | de |
| dc.type.publicationtype | PhDThesis | de |
| dcterms.accessRights | open access | |
| eldorado.secondarypublication | false | de |