Simulation real time

The engine.
Post Reply
kyc212
Posts: 4
Joined: Thu Apr 13, 2017 12:59 pm

Simulation real time

Post by kyc212 »

Hi,
I'm working on a project which involves connecting to a motion tracker and reading position and orientation data from it in real-time. I've just built a DBN model to predict object transition at various points of time. My model is DBN, but I need some clarification/help in few things.
Since the object can change position/location at two or more seconds (e.i from Square A to Square B), is it a good approach to have time slices for every second? I don't think it is!!!!- Genie only allows up to 1000.

One possible solution to my problem is to have a fixed unrolled network (lets say from observation my object only visits 60 squares per day) so my unrolled network could have only t=60. Looping the DBN and set evidence at each time slice for each transition seems like an obvious solution, but how can I export the evidence in my model.xdsl. I have to create a report of all visits and time for each object.

Apologies if the question is poorly phrased,

Any advice or references would be much appreciated

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

Re: Simulation real time

Post by shooltz[BayesFusion] »

how can I export the evidence in my model.xdsl.
The xdsl format supports saving evidence as cases. This functionality is only available from core C++ library (not from jSMILE or Smile.NET).
I have to create a report of all visits and time for each object.
You'll need to write a program to create such report.
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Simulation real time

Post by marek [BayesFusion] »

Hi,

GeNIe's limitation is just for your own sake :-) -- 1000 slices is possibly a computational challenge but it could work, depending on the complexity of your network. Coming back to your main question whether you should model only slices during which something happens or all slices, it depends on the nature of the problem. If the only thing that matters for predicting the future is the state of your network, I would say that you don't need to model the idle cycles. This will make your network much more efficient computationally. In fact, if you model just a handful of events, it will be very easy. Just watch the interpretation of your time slices -- they are no longer linear seconds but rather events. Moving the time scale along with your object, i.e., treating the first slice as time zero each time, sounds also like a good idea.
Does this help? Please ask further if it does not!
Cheers,

Marek
Post Reply