Does Smile use something like AIS-BN?

The engine.
Post Reply
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

Does Smile use something like AIS-BN?

Post by Jochen0x90h »

Hi!

I found the paper AIS-BN: An Adaptive Importance Sampling Algorithm for Evidential Reasoning in Large Bayesian Networs by Jian Cheng and Marek J. Druzdzel
http://www.jair.org/media/764/live-764-1924-jair.pdf. This paper originated in the University of Pittsburgh, therefore it would be interesting if SMILE is using this or an improved version of this algorithm. This is important when evaluating large models such as the QMR-DT network, CPCS, ANDES or other networks that are not really computable analytically.

Independent form this question, can you recommend using SMILE on large scale networks?
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Does Smile use something like AIS-BN?

Post by marek [BayesFusion] »

Thanks for your interest in AIS-BN and SMILE. Yes, SMILE implements AIS-BN. When you use GeNIe, you can select the algorithm to be used in inference by means of the Network->Algorithm menu. AIS Sampling is what you want. For very large networks, for which the exact (Clustering) algorithm does not work, please try Relevance-based Decomposition, also an exact algorithm -- it will often update very large networks. If you are dealing with networks that SMILE really cannot handle exactly (there are not so many of them 8) ), I recommend EPIS Sampling -- it is a significant improvement on the AIS-BN algorithm. I believe that all of the networks that you listed can be computed exactly using SMILE/GeNIe. Our software is really fast and efficient 8).
I hope this helps.

Marek
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Does Smile use something like AIS-BN?

Post by shooltz[BayesFusion] »

marek wrote:For very large networks, for which the exact (Clustering) algorithm does not work, please try Relevance-based Decomposition, also an exact algorithm -- it will often update very large networks
Just a side note: clustering will switch to relevance-based decomposition automatically if needed.

I'd definitely recommend marking nodes of interest as targets if not all posteriors are required. This may greatly reduce the memory pressure.
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

Re: Does Smile use something like AIS-BN?

Post by Jochen0x90h »

Thanks for your reply. How do you ensure that the loopy belief propagation step of EPIS-BN converges to an approximate solution and does not start to oscillate? Is it always better than nothing even if it starts to oscillate?

I'm on a mac so I will try wine first to get the gui working.

-Jochen
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Does Smile use something like AIS-BN?

Post by shooltz[BayesFusion] »

Jochen0x90h wrote:I'm on a mac so I will try wine first to get the gui working.
If wine doesn't work for you, consider using VirtualBox or VMWare with one of Microsoft's free IE dev VMs. I'm using a Mac with Windows VM currently and it works quite nicely.

http://www.modern.ie/en-us/virtualization-tools
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Does Smile use something like AIS-BN?

Post by marek [BayesFusion] »

Jochen0x90h wrote:How do you ensure that the loopy belief propagation step of EPIS-BN converges to an approximate solution and does not start to oscillate? Is it always better than nothing even if it starts to oscillate?
We run LBP for only a few steps (I forgot how many but I believe it is five or so), so there is not even opportunity for it to oscilate 8). The idea is to start sampling from a distribution that is close to the posterior. It turns out that the five steps we do gets us really close to where we want to be. The rest is done by sampling.
I hope this helps,

Marek
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

Re: Does Smile use something like AIS-BN?

Post by Jochen0x90h »

Thanks for the info.
I'd like to try the java binding. It would be convenient to have it in a maven repository (see http://search.maven.org and http://maven.apache.org/guides/mini/gui ... pload.html). Also i would suggest changing the package name from smile to either edu.pitt.sis.smile or edu.pitt.ischool.smile or only edu.pitt.smile.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Does Smile use something like AIS-BN?

Post by shooltz[BayesFusion] »

Jochen0x90h wrote:Thanks for the info.
I'd like to try the java binding. It would be convenient to have it in a maven repository (see http://search.maven.org and http://maven.apache.org/guides/mini/gui ... pload.html).
As I write this, we're configuring the artifactory on one of our servers (this is related to one of the projects we're in). You'll be able to add the additional repo URL to your pom file and have jSMILE pulled in.
Also i would suggest changing the package name from smile to either edu.pitt.sis.smile or edu.pitt.ischool.smile or only edu.pitt.smile.
Unlikely due to backward compat reasons :)
Jochen0x90h
Posts: 19
Joined: Thu Jul 11, 2013 1:20 pm

Re: Does Smile use something like AIS-BN?

Post by Jochen0x90h »

Is the URL available by now?
Post Reply