Re: questions about PhaseIIIa

Subject:

Comments:
: Julie, : Generally, your understanding is correct. : The core implementation of phaseIIIa is an approximating BBN inference algorithm. : -Input: <1> an .xml XMLBIF file representing a BBN : <2> a .dat file containing k cases data points for that BBN. This file can be generated by the data generator. : Followed is an example .dat file with 1 case: : *******Asia.dat******* : 0 1 0 1 1 0 0 0 ---------> 1 = evidence node, 0 = query node : A B C D E F G H ---------> node names : a1 b1 c2 d1 e2 f1 g1 h1 ---------> node values : ********************** : -Procedure and Output: : 1. The input module loads .xml into an AllNodes object in memory, and loads .dat file into a suitable object in : memory. : 2. The core implementation of your computing module computes P(X|E), the probabilities of query nodes given : evidence, for each case in the .dat file . Here X={A,C,F,G,H}, E={B,D,E}. : 3. Your computing module loops through all possible values that each query node can take, computes P(X|E) = : P(X,E)/P(E) according to the Bounded-variance algorithm , and then choose the value corresponding to the maximal : probability as the prediction. : 4. It then counts each error (where an error is a difference between the prediction and the actual value in the : line in the .dat file), and calculates the inferential loss. : 5. Do 2,3,4 for all cases in .dat file, sum up all inferential loss , output the final value or compute the MSE : procedure Bounded-variance (see P15 in Dagum&Luby's paper) : Input: relative error 0<e<2, failure probability 0<delta<1, a set of k instantiated nodes W = {W1, W2, ..., Wk} : Output: P(W) : Begin : Compute S* and A(the product of u1, u2, ... uk) from e and delta; : S=0; : t=0; // t - the number of sampling experiments : While S<S* do { : Generate_Sample z; : compute new S according to z; : t++; : } : T = t; : Output P(W) = A(S*)/T : End : Procedure Generate_Sample : Input: a BBN, a set of evidence nodes E : Output: a sample z for all nodes : Begin : Order the BBN nodes so that parent nodes occur before child nodes; : Generate value for all nodes starting from root nodes, if a root node belongs to E then instantiate it to its : observed value; If not, we generate a value for this node according to its prior probability; : Once all root nodes are instantiated, proceed to the set of these nodes that have all parents instantiated. Again, : if it belongs to E then instantiate it to its observed value; If not, we generate a value for this node according to its : parents' values and its PDT; : We proceed until all nodes in the BBN are instantiated; : Output the sample instantiation z; : End : : Note: The Input module and Generate_Sample Procedure has been implemented in Data generator, but it may need some minor : modifications. : : -hpguo : : : -In a given phase, I take in a case from a line in the .dat file. Whichever : : values it has for my evidence nodes I hold constant, and then try to predict : : the outcome of all other (query) nodes. : : -To predict the values of these query nodes, I loop through all possible : : values that each query node can take, evaluate the probability of each : : possible value based on its parents, and then choose the value corresponding : : to the maximal probability as my prediction. I continue this until all : : query nodes have been predicted. : : -I then count each error (where an error is a difference between my : : prediction and the actual value in the line in the .dat file), and calculate : : the inferential loss. : : If my above ideas are correct, I have the following questions on how to : : implement them: : : -Where does the dataGenerate function come in from DataGen? It seems to me : : that if I randomly generate all query root nodes, then I would then be able : : to infer all other query nodes by simply taking the max probability : : associated with the given parents. Or does the dataGenerate generate : : appropriate values while taking the parent's values into account? : : -How do I access the probability table? I know that each node has a Vector : : called probabilitytable associated with it, but how do I tell from this : : which values correspond to a certain instantiation of the node's parents? : : -I believe I'm supposed to randomly generate (using dataGenerate) values for : : the query nodes over and over until a certain value is "honed in on". Does : : this correspond to the place in Dagum and Luby's algorithm where they loop : : till the sum over all instantiations of the product of probabilites that : : that instantiation will have given its parents until that sum is greater : : than the specified S*? : : If at all possible, I'd really like the method and class specifications that : : you mentioned as well as a somewhat layman's explanation as to how my module : : basically operates, and how exactly it makes use of all the classes and : : methods from the data generator. : : I'm sorry I keep having various confusions pop up, but I want to make sure : : that I fully understand what I am trying to do and how best to go about it : : before attacking things any more. I feel like I have a feel for the basic : : ideas behind the module, but am still unclear on the more detailed : : specifications of implementation. I'd really like to get this module : : finished this week, so your comments are greatly appreciated.

Optional Link URL:
Link Title:
Optional Image URL: