What is the best way to handle inference in networks containing equation nodes where evidence has been assigned to one of the child nodes?
I have attached a simple example to illustrate this. In the example network, whenever I set evidence to Node 2 or 4 and try to update the beliefs, GeNIe will run into an infinite loop or crash. If I only assign evidence to the root nodes, it performs the update without problems.I have tried different algorithms provided by GeNIe, but couldn't get it to work. I have also looked into discretizing the nodes, but this does not look as a viable option for the size of the network I am using in my research.
Thanks,
Adrian
Inference with Equation Nodes
Inference with Equation Nodes
- Attachments
-
- ContinuousNetwork.xdsl
- (1.6 KiB) Downloaded 517 times
-
- Site Admin
- Posts: 1417
- Joined: Mon Nov 26, 2007 5:51 pm
Re: Inference with Equation Nodes
SMILE support for evidence in continuous networks is based on a research-quality code and is not very robust if you set evidence to nodes with parents. You may consider using network discretization to succesfully perform inference (in GeNIe, enable Network|Algorithm|Autodiscretize and specify discretization intervals in node properties).aambrus wrote:What is the best way to handle inference in networks containing equation nodes where evidence has been assigned to one of the child nodes?
Re: Inference with Equation Nodes
Thanks for the reply. Does jSMILE contain methods for enabling auto-discretization and setting the discretization intervals on a continuous node?
Adrian
Adrian