Good presentation! Did you perform any dimensionality reduction (e.g. PCA, stepwise regression, clustering) or how did you deal with such big number of variables?
Thank you for your question.
For both networks, we did first a reduction on characteristics that are definitely not important for the tool selection or the quotation costing. For example, the colour code is added after the grinding takes place and has the same cost for all milling cutters.
This is how we ended up with more than 50 input characteristics. That sound a lot in the first place but many characteristics are nominal and sometimes empty for some of the milling cutters. This reduces the complexity by a small degree.
We didn’t perform any more reduction on this input data because our network was not rated on the time performance and only on the result.
We saw early in the research that the network will weigh some characteristics with a bad score after only a few iterations. I then dropped the characteristics but this only improved the training time and had no impact on the result. Together with the company, we decided to leave every variable in the input data.
For a better understanding, we performed a PCA for both problems but this is not part of our network which I presented. It was more for the company to understand the main characteristics and to help them to get confident in using the network.
Good presentation! Did you perform any dimensionality reduction (e.g. PCA, stepwise regression, clustering) or how did you deal with such big number of variables?
Thank you for your question.
For both networks, we did first a reduction on characteristics that are definitely not important for the tool selection or the quotation costing. For example, the colour code is added after the grinding takes place and has the same cost for all milling cutters.
This is how we ended up with more than 50 input characteristics. That sound a lot in the first place but many characteristics are nominal and sometimes empty for some of the milling cutters. This reduces the complexity by a small degree.
We didn’t perform any more reduction on this input data because our network was not rated on the time performance and only on the result.
We saw early in the research that the network will weigh some characteristics with a bad score after only a few iterations. I then dropped the characteristics but this only improved the training time and had no impact on the result. Together with the company, we decided to leave every variable in the input data.
For a better understanding, we performed a PCA for both problems but this is not part of our network which I presented. It was more for the company to understand the main characteristics and to help them to get confident in using the network.