I was wondering, whats main idea behind this competition is?
\n\nLet's say we have build a classifier who predicts manufacturing failures with a high accuracy.\nCan I use this classifier to do simulations with varying parameters of my production line to see if I produce less failures?\nI guess the usual simulation tools for a manufacturing process use some 'optimal' calculated parameters , but cannot account for the variations and the unknown physical connections in the parameters. \nSo is it possible to use a classifier for this multiparameter optimization??
\n\nIs something like this already done? Are there papers about it?
", "messages": [ { "id": "140948", "postDate": "10/24/2016 10:20:48", "content": "I was wondering, whats main idea behind this competition is?
\n\nLet's say we have build a classifier who predicts manufacturing failures with a high accuracy.\nCan I use this classifier to do simulations with varying parameters of my production line to see if I produce less failures?\nI guess the usual simulation tools for a manufacturing process use some 'optimal' calculated parameters , but cannot account for the variations and the unknown physical connections in the parameters. \nSo is it possible to use a classifier for this multiparameter optimization??
\n\nIs something like this already done? Are there papers about it?
", "rawMarkdown": "I was wondering, whats main idea behind this competition is?\r\n\r\nLet's say we have build a classifier who predicts manufacturing failures with a high accuracy.\r\nCan I use this classifier to do simulations with varying parameters of my production line to see if I produce less failures?\r\nI guess the usual simulation tools for a manufacturing process use some 'optimal' calculated parameters , but cannot account for the variations and the unknown physical connections in the parameters. \r\nSo is it possible to use a classifier for this multiparameter optimization??\r\n\r\nIs something like this already done? Are there papers about it?", "votes": null }, { "id": "140990", "postDate": "10/24/2016 15:17:13", "content": "The main idea is to identify sources of production failures so that they may corrected and hopefully save the company time and money. Such failures could result from a lot of places such as a particular machine, a particular production path, a particular product, particular time of day, etc. This information is what they are hoping to learn.
\n\nAs far as your comment about the "usual simulation tools" for manufacturing processes, I think you may be overestimating the complexity of what is currently available. In reality, the participants in this competition are the ones building the tools you're referring to.
", "rawMarkdown": "The main idea is to identify sources of production failures so that they may corrected and hopefully save the company time and money. Such failures could result from a lot of places such as a particular machine, a particular production path, a particular product, particular time of day, etc. This information is what they are hoping to learn. \r\n \r\nAs far as your comment about the \"usual simulation tools\" for manufacturing processes, I think you may be overestimating the complexity of what is currently available. In reality, the participants in this competition are the ones building the tools you're referring to.", "votes": null }, { "id": "140997", "postDate": "10/24/2016 16:15:23", "content": "But if I train a black box model like a neural network, how can I know what the sources of the production failures are?
\n\nfrom the xgboost models i can at least look at the feature importance.
", "rawMarkdown": "But if I train a black box model like a neural network, how can I know what the sources of the production failures are?\r\n\r\nfrom the xgboost models i can at least look at the feature importance.", "votes": null }, { "id": "141000", "postDate": "10/24/2016 16:22:32", "content": "If you choose to use a black-box model, you give up the ability to look at the causes, you're just wanting a way to predict failures. If you care about feature importance you must, as you alluded, choose something else. It depends on what your goal is.
\n\nIn this case, I would expect identifying the causes is important, so that they can be fixed for the long-term. But there are still many applications for which a black-box model is fine. For a hypothetical example that still uses this data, if it turned out a neural network was by far the best model, the parameters for every part to be made could be passed through the network before the part was produced, and those that were flagged as failures could be stopped before they ever start. Kind of a contrived example, but hopefully you get my point.
", "rawMarkdown": "If you choose to use a black-box model, you give up the ability to look at the causes, you're just wanting a way to predict failures. If you care about feature importance you must, as you alluded, choose something else. It depends on what your goal is. \r\n\r\nIn this case, I would expect identifying the causes is important, so that they can be fixed for the long-term. But there are still many applications for which a black-box model is fine. For a hypothetical example that still uses this data, if it turned out a neural network was by far the best model, the parameters for every part to be made could be passed through the network before the part was produced, and those that were flagged as failures could be stopped before they ever start. Kind of a contrived example, but hopefully you get my point.", "votes": null }, { "id": "141071", "postDate": "10/25/2016 01:37:43", "content": "Large manufacturers have been solving these problems for decades using tools that are decades old. Sometimes it is hard to overcome inertia, especially when you already have a tool that works well. (Hint, I'm sure that Bosch knows all about the Id sequence issue. I see a similar phenomenon in my factory. They have tools to find those kinds of correlations. They probably don't use complicated ensembles like top competitors.)
\n\nI think this competition is interesting because people are trying to solve it with many new techniques. As Alan mentioned, there are several ways that manufacturers might use these models. However, I think the real value is in learning new ways to approach the problem.
", "rawMarkdown": "Large manufacturers have been solving these problems for decades using tools that are decades old. Sometimes it is hard to overcome inertia, especially when you already have a tool that works well. (Hint, I'm sure that Bosch knows all about the Id sequence issue. I see a similar phenomenon in my factory. They have tools to find those kinds of correlations. They probably don't use complicated ensembles like top competitors.)\r\n\r\nI think this competition is interesting because people are trying to solve it with many new techniques. As Alan mentioned, there are several ways that manufacturers might use these models. However, I think the real value is in learning new ways to approach the problem.", "votes": null } ], "comments": [ { "id": 140990, "author_name": "apryor6", "author_url": "", "post_date": "10/24/2016 15:17:13", "content": "The main idea is to identify sources of production failures so that they may corrected and hopefully save the company time and money. Such failures could result from a lot of places such as a particular machine, a particular production path, a particular product, particular time of day, etc. This information is what they are hoping to learn.
\n\nAs far as your comment about the "usual simulation tools" for manufacturing processes, I think you may be overestimating the complexity of what is currently available. In reality, the participants in this competition are the ones building the tools you're referring to.
", "votes": null, "replies": [] }, { "id": 140997, "author_name": "weterp", "author_url": "", "post_date": "10/24/2016 16:15:23", "content": "But if I train a black box model like a neural network, how can I know what the sources of the production failures are?
\n\nfrom the xgboost models i can at least look at the feature importance.
", "votes": null, "replies": [] }, { "id": 141000, "author_name": "apryor6", "author_url": "", "post_date": "10/24/2016 16:22:32", "content": "If you choose to use a black-box model, you give up the ability to look at the causes, you're just wanting a way to predict failures. If you care about feature importance you must, as you alluded, choose something else. It depends on what your goal is.
\n\nIn this case, I would expect identifying the causes is important, so that they can be fixed for the long-term. But there are still many applications for which a black-box model is fine. For a hypothetical example that still uses this data, if it turned out a neural network was by far the best model, the parameters for every part to be made could be passed through the network before the part was produced, and those that were flagged as failures could be stopped before they ever start. Kind of a contrived example, but hopefully you get my point.
", "votes": null, "replies": [] }, { "id": 141071, "author_name": "jeffhebert", "author_url": "", "post_date": "10/25/2016 01:37:43", "content": "Large manufacturers have been solving these problems for decades using tools that are decades old. Sometimes it is hard to overcome inertia, especially when you already have a tool that works well. (Hint, I'm sure that Bosch knows all about the Id sequence issue. I see a similar phenomenon in my factory. They have tools to find those kinds of correlations. They probably don't use complicated ensembles like top competitors.)
\n\nI think this competition is interesting because people are trying to solve it with many new techniques. As Alan mentioned, there are several ways that manufacturers might use these models. However, I think the real value is in learning new ways to approach the problem.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "140948": "I was wondering, whats main idea behind this competition is?\r\n\r\nLet's say we have build a classifier who predicts manufacturing failures with a high accuracy.\r\nCan I use this classifier to do simulations with varying parameters of my production line to see if I produce less failures?\r\nI guess the usual simulation tools for a manufacturing process use some 'optimal' calculated parameters , but cannot account for the variations and the unknown physical connections in the parameters. \r\nSo is it possible to use a classifier for this multiparameter optimization??\r\n\r\nIs something like this already done? Are there papers about it?", "140990": "The main idea is to identify sources of production failures so that they may corrected and hopefully save the company time and money. Such failures could result from a lot of places such as a particular machine, a particular production path, a particular product, particular time of day, etc. This information is what they are hoping to learn. \r\n \r\nAs far as your comment about the \"usual simulation tools\" for manufacturing processes, I think you may be overestimating the complexity of what is currently available. In reality, the participants in this competition are the ones building the tools you're referring to.", "140997": "But if I train a black box model like a neural network, how can I know what the sources of the production failures are?\r\n\r\nfrom the xgboost models i can at least look at the feature importance.", "141000": "If you choose to use a black-box model, you give up the ability to look at the causes, you're just wanting a way to predict failures. If you care about feature importance you must, as you alluded, choose something else. It depends on what your goal is. \r\n\r\nIn this case, I would expect identifying the causes is important, so that they can be fixed for the long-term. But there are still many applications for which a black-box model is fine. For a hypothetical example that still uses this data, if it turned out a neural network was by far the best model, the parameters for every part to be made could be passed through the network before the part was produced, and those that were flagged as failures could be stopped before they ever start. Kind of a contrived example, but hopefully you get my point.", "141071": "Large manufacturers have been solving these problems for decades using tools that are decades old. Sometimes it is hard to overcome inertia, especially when you already have a tool that works well. (Hint, I'm sure that Bosch knows all about the Id sequence issue. I see a similar phenomenon in my factory. They have tools to find those kinds of correlations. They probably don't use complicated ensembles like top competitors.)\r\n\r\nI think this competition is interesting because people are trying to solve it with many new techniques. As Alan mentioned, there are several ways that manufacturers might use these models. However, I think the real value is in learning new ways to approach the problem." }, "source": "meta" }