Prediction of a Value.

The front end.
Post Reply
ramzi9572
Posts: 6
Joined: Mon Apr 26, 2021 10:24 am

Prediction of a Value.

Post by ramzi9572 »

Hi everyone,

I'm new user of "Genie" and i liked the software sofar. But I can't get the solution of my case. So I try now to describe it and hope that i get it solved.

I would like to get a Prediction of a value for the product throughput "lead time" time in a manufacturing line. So I need to use the BN as a Regression model. I dont know if i used it right. I've tried a lot but nothing has helped. I have 4 machines and I measured different times for each one. Unfortunately, I cannot get my dataset in the right form. Here is my dataset for the first 3 machines. In columns F and M you can see the processing times. the times B to E are the times that are necessary for F and Column G-L for M etc..
So my questions are:
Q1: I created my BN manually and then i learnd the parameter but after the learning i still have just 2 states (0 and 1) shall i add the other states manually? For example i have the "entry times" and the are 100 times. Shall i add the other 98 states manually? I thought tha will be done after the randomize with the EM.

Q2: Why after 100 states for a paramter does my GeNie crash? Do i have a max number of states? I have a good Memory btw.

Q3: I created the network but i dont know if it's right and for my case. I can set a Evidence and get a probality for the next lead times. I dont think that i need a DBN right?

Q4: I dont know how to get the right form of my Dataset. can someone please take a look? My targets are DLZ_M1/DLZ_P1/DLZ_M2/DLZ_M3.

Thanks a lot
Attachments
GeNie-4.csv
(19.35 KiB) Downloaded 157 times
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Prediction of a Value.

Post by marek [BayesFusion] »

Hi Ramzi,

Let me try to answer your questions. Thanks for attaching the data file -- it helps. The model would be very helpful but since I don't have it, I will try to base my responses on the data alone. If I am off, please attach the model as well.

The format for time is something that GeNIe will not recognize. Your data should be either continuous and numerical or categorical. Numerical means that the values should be numbers (integer or floating point). One thing that you can do is perhaps convert values like "12:38,6566" to numbers that express the number of seconds, for example. Another problem with your data is that you use comma for decimal point. The Anglo-Saxon convention is that we use decimal point for that. This should be an easy change in Excel. I have noticed also that your variable names contain special characters. While GeNIe will swallow that, it will be easier on you to have variable/column names take the syntax of variables in a programming language. GeNIe turns them into IDs, which are pretty much variable names. Once you convert your data to the format that GeNIe recognizes, things will go smoothly.

Q1: The two states are probably what GeNIe made of your numerical values. This should change once you change the format of your data. I assume your variables are discrete. You should discretize your data and give the intervals exactly the names of states of your corresponding variables.

Q2: Quite possibly, variables in your model turned out to have many, many states. When you have many states, you will exhaust your memory, no matter how large it is -- please note that CPTs grow exponentially with the number of parents and the base is equal to the number of states.

Q3: Hard to say if you need a DBN, as I don't know your problem well but from what you described, it seems that a static network will be just fine.

Q4: I described the problems in your data set format above. Please try to polish your data.

I hope this helps,

Marek
ramzi9572
Posts: 6
Joined: Mon Apr 26, 2021 10:24 am

Re: Prediction of a Value.

Post by ramzi9572 »

Thank you Marek for your response.

I have already corrected my dataset and now it works. I have 2 Q.

Q1: All my variabls are continuous and I need a con. prediction. I did a Test-Only validation. With corss-validation i became a error, so i stay by Test only. How can i get the right cont. prediction?

Q2: If a variable is constant. It delivers no Info for my target, so i delet the connection for this node. Is this way right?

Thank alot
ramzi9572
Posts: 6
Joined: Mon Apr 26, 2021 10:24 am

Re: Prediction of a Value.

Post by ramzi9572 »

UPdate: I know why i get error. I had a node as desicion node and i changed it to general. Then the problem is solved.

In all arts of the Validation. I become about 72% Accuracy. is this a good value? How can i improve it?
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Prediction of a Value.

Post by marek [BayesFusion] »

Q1: All my variabls are continuous and I need a con. prediction. I did a Test-Only validation. With corss-validation i became a error, so i stay by Test only. How can i get the right cont. prediction?
I'm afraid it is not possible in GeNIe yet. We will get to it quite likely next year.
Q2: If a variable is constant. It delivers no Info for my target, so i delet the connection for this node. Is this way right?
Yes, a variable that is a constant brings nothing useful to the model. Unless you want to keep it for clarity of the model, you can can as well delete it.
In all arts of the Validation. I become about 72% Accuracy. is this a good value? How can i improve it?
There is no way I can tell. Some problems are so hard that the best you can do is being a little better than random guessing. Some problems are easy and you can reach accuracy close to 100%. I would compare your model's accuracy to the results that other people are publishing.

I hope this helps,

Marek
ramzi9572
Posts: 6
Joined: Mon Apr 26, 2021 10:24 am

Re: Prediction of a Value.

Post by ramzi9572 »

Thanks alot Marek. Alle Questions are answered. Appreciated.

for the con. prediction. I decided to do it with the package BNLearn from R. it works. I hope I can help the other with this info.
Post Reply