ENBIS-16 in Sheffield

11 – 15 September 2016; Sheffield Abstract submission: 20 March – 4 July 2016

Evolutionary Operation for Contemporary Processes

14 September 2016, 09:20 – 09:40


Submitted by
Bart De Ketelaere
Bart De Ketelaere (Catholic University of Leuven), Mohammad Rohail Taimour (Catholic University of Leuven), Peter Goos (Catholic University of Leuven)
Evolutionary Operation (EVOP) was first mentioned by George Box in 1957 as a simple method that can be used to improve full scale processes while they are running. The rules associated with it allowed EVOP to be calculated by hand on a sheet of paper so that process operators could implement the adapted settings of the process and evaluate the adapted process. This idea of sequentially perturbing the process and to learn from it is both simple and effective, but it never became a standard operation in practice. The main reason is probably that EVOP was, until recently, never studied in more detail for processes that are common in current practice: they are governed by a multitude of parameters and sampling intervals are very short requiring fast and automatic computation.
Rutten and co-workers (2014) understood this gap and performed research into EVOP schemes that are suitable for higher dimensional processes and automatic execution. In essence, they proposed several rules for implementing EVOP for low and high dimensions, and applied them in simulation studies as well as in real-life cases. One of the main open issues that they discussed is the choice of base design for EVOP when the dimensionality prohibits the standard full factorial to be used. In such cases, smaller base designs are required, but it is unclear what those base designs should look like to obtain an optimal EVOP scheme. Optimal here means that the process improvement is achieved using a minimal amount of experimental effort. Preliminary simulations performed by Rutten (2014) pointed in the direction of favoring low power designs. The aim of this contribution is to refine the simulations to determine the optimal properties for the base designs in EVOP, for low as well as for high dimensions.
The approach taken starts from a simple numerical simulation study where a linear process needs to be optimized. We initially use an analytical approach to this simulation yielding more accurate results than those obtained by Rutten (2014). The main conclusions from these studies will be discussed and can be summarized as follows: (1) base designs with a low power outperform larger designs having a higher power; (2) base designs for higher dimensional processes ask for a higher power when compared to lower dimensions; (3) replication of the base design when no process improvement direction is found is advantageous for base designs with very low power, but not for base designs with higher power.

Return to programme