Suggestion on "network structure too complex"

The front end.
Post Reply
emavro
Posts: 5
Joined: Wed Aug 01, 2018 5:48 pm

Suggestion on "network structure too complex"

Post by emavro » Thu Oct 18, 2018 12:08 pm

Hello,
I am working on a model for risk calculation.
It ends at 60 nodes of partial risk (R01 up to R60) in groups of 10's for each of the 6 syb-systems. Each node R1-60 has 10 states.
I have managed to calculate each sub-system's risk RS1,RS2,RS3, RS4, RS5 and RS6 by parent divorcing but when I try further aggregation the output is always "Error (-42): UpdateBeliefs failed - network structure is too complex."
Could you propose an alternative solution for total risk calculation?

Thank you.

shooltz[BayesFusion]
Site Admin
Posts: 1181
Joined: Mon Nov 26, 2007 5:51 pm

Re: Suggestion on "network structure too complex"

Post by shooltz[BayesFusion] » Thu Oct 18, 2018 5:23 pm

Are you going to use the calculated posteriors for all nodes in your model, or only a subset? For the latter case, try marking the nodes you're interested in as targets. This reduces the complexity of calculations GeNIe has to perform.

emavro
Posts: 5
Joined: Wed Aug 01, 2018 5:48 pm

Re: Suggestion on "network structure too complex"

Post by emavro » Sat Oct 20, 2018 5:56 am

Thanks for the reply,
Only on a subset. I have already marked as "target" the nodes that I am interested for. But the message keeps the same about too complex network.
Do you have any other ideas?

shooltz[BayesFusion]
Site Admin
Posts: 1181
Joined: Mon Nov 26, 2007 5:51 pm

Re: Suggestion on "network structure too complex"

Post by shooltz[BayesFusion] » Sat Oct 20, 2018 11:01 am

Do you have any other ideas?
Not really. We'll have significant performance improvements for network containing noisy nodes in the upcoming GeNIe release. If your network contains such nodes, it may be worth trying.

emavro
Posts: 5
Joined: Wed Aug 01, 2018 5:48 pm

Re: Suggestion on "network structure too complex"

Post by emavro » Sat Oct 20, 2018 3:46 pm

Thanks,
I will definitely give it a try. But, unfortunately, for the current version of my model (due to deadline restrictions), I might use the 6 subsystem partial risk nodes RS1 to RS6, copy-pasted to a new model just for the final aggregation procedure for the calculation of the Total Risk (Rtot) node.

By the way, is there any documentation on the maximum size of the tables that the program can handle? Because I couldn't find it on Genie 's manual. If my little statistics knowledge is correct, for my model's final node, node Rtot, I should need a table with 10^61 cells.
Thank you for your time!

shooltz[BayesFusion]
Site Admin
Posts: 1181
Joined: Mon Nov 26, 2007 5:51 pm

Re: Suggestion on "network structure too complex"

Post by shooltz[BayesFusion] » Mon Oct 22, 2018 7:45 pm

By the way, is there any documentation on the maximum size of the tables that the program can handle?
The program does not have any built-in limitations on the CPT size. Your RAM is the deciding factor.

Note however, that in addition to the memory used by the CPTs, you'll need extra memory to calculate the posteriors. The amount required depends on the structure of the network, currently selected evidence and the target set.
I should need a table with 10^61 cells.
That's rather above the capabilties of the hardware which you can buy off-the-shelf :)

emavro
Posts: 5
Joined: Wed Aug 01, 2018 5:48 pm

Re: Suggestion on "network structure too complex"

Post by emavro » Tue Oct 23, 2018 4:11 am

Thanks for your informative answer!

Post Reply