Temporal Bayesian Networks

[nutterb]

• Time 0 model (baseline conditions and time-dependent variables at T=1)
  • Time t model (maps time-dependent variables at time t to the same at time t+1) • unroll.markovNetwork(startTime=NULL, stoptime)
• This function just creates a pgmNetwork object from the Markov model by repeating the transition structure over a user-specified span of time o if startTime==NULL, unroll the network from t=0 to t=stoptime
  • prior distributions defined by parameterized densities for the t=0 nodes o if startime>=1, unroll the network from time=startTime to time=stopTime
  • We may need to assume all (root) nodes at time=startTime have either been observed or computed as the posterior mode from a previous run of JAGS

[jarrod-dalton]

Here is an excellent paper that is worth understanding. They implement a Bayesian non-linear state space model using a small dataset and WinBUGS. I think this could be represented as a temporal Bayesian network, composed of two subnetwork (HydeNet) model objects - the one mapping time 0 to time 1, and the one mapping time t to time t+1.

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.365.9015&rep=rep1&type=pdf

In case the link doesn't work, the article is:

Meyer, Renate, and Russell B. Millar. "BUGS in Bayesian stock assessments." Canadian Journal of Fisheries and Aquatic Sciences 56.6 (1999): 1078-1087.

[jarrod-dalton]

After reflecting on that paper, I'm getting a better idea of how temporal models with HydeNet should look. Instead of writing about it here, I'll start a vignette.