Feedback
Hi Ricky, Thank you so much for your feedback! 1. Yes indeed I could add a link to the page for Bayesian network to make things clearer. 2. Well the traditional approach is to take the entire JPD and condition and marginalize on it. However the space and time complexity for this is O(dn) where 'd' is the size of the largest domain and 'n' is number of variables.n We do feel the need to normalize at the very end even though the factors don't need to sum to 1. But after multiplying factors we need to normalize (as shown in the demo). In essence, 'summing out a variable' kind of relates to marginalizing and 'assigning a variable' relates to conditioning intuitively. I'm not sure what you mean by the formula part of the question? 3. As long as the conditional independencies are not violated (i.e.all the necessary dependencies are represented through the links between nodes in the Bayesian network), any elimination order should be valid. The maximum factor size could however be a concern. 4. Sorry, I meant to say 'n' binary variables. I mentioned binary because usually variables have few parents in Bayesian network. Thank you for you comments. Regards, Ritika