Dr. Amos Ng, WP leader
Dr. Alexander Karlsson
Dr. Bo Stenberg (Agroväst)
MSc Leif Pehrsson (VCC)
MSc Florian Siegmund
Research Question
The following research questions are addressed in WP7:
- Uncertainty representation: How errors/uncertainties in model results are coupled to errors/uncertainties in input parameters?
- How robust Pareto-optimal solutions can be effectively sought when there exist different sources of uncertainty/imprecision in the optimization process.
- How can one effectively estimate and visualize the effects when uncontrollable and/or unpredictable parameters vary from a model/simulation which subject to different sources of uncertainty?
- Can the same method/toolset be used for discrete-event simulation models for manufacturing and wheat growth analytical model used in precision agriculture?
Relevance to UMIF
WP7 is related to uncertainty representation in parameters and uncertainty visualization in model outputs (UMM research area 1 and 2). Different uncertainty representations (connection to UMM area) will be tested and a graphical model such as Bayesian network (link to Impact Analysis, Area 3) will be investigated.
Collaboration with Industry
Volvo Car Corporation (VCC) provides the problem formulations and data for the manufacturing application area, where uncertainty may be caused by unforeseen events like machine breakdown, but more often the uncertainty is related to changes in product design or demand. Decision making in this application area will involve the design, improvement and analysis of new/existing production lines/facilities. Agroväst will contribute to the use of wheat growth model (SIRIUS). The developed toolset can also be tested with the crop growth data generated in SIRIUS for investigating the effect of the variations in unpredictable and uncontrollable parameters and perform risk analysis for the decision-making support of farmers.
Approach
WP7 investigates the algorithmic aspects of evolutionary multi-objective optimization (EMO) algorithm in handling uncertainty. Specifically a dynamic resampling approach using Bayesian statistics and other heuristics to optimally allocate simulation budgets is being investigated. Other deliverables include:
- Robustness analysis and visualization techniques implemented in existing toolset for manufacturing simulation and optimization.
- Risk analysis prototype toolset for model-based wheat growth prediction.
WP Results & Status
A reference-point guided EMO algorithm with dynamic resampling is being developed and under testing. It is being tested on a cost optimization problem provided by VCC. In general, its performance and efficiency outperforms significantly EMO with static resampling but benchmarking with other dynamic resampling strategies needs more experiments. A prototype visualization tool is being developed to visualize the dynamic allocation of the algorithm as well as other uncertainty information of the solutions generated from the optimization. VCC has also completed a simulation model ready to be tested with robustness analysis. Two project meetings/workshops with Agroväst and University of Uppsala were conducted but the case study in precision agriculture is yet to be defined.
Related Work
WP7 is related to other simulation-based optimization research at the university, in particular to the VINNOVA FFI-HSO project.