/* * Recover the maximizing variables going back through the * maximizing bucket_tree; the variables are returned as an array * of markers (non-explanation variables get INVALID_INDEX). */ private int[] backward_maximization() { int i, j; int bi = bucket_tree.length - 1; DiscreteFunction back_df; Bucket b = bucket_tree[bi]; // If there are no explanation variables in the BayesNet, return null if (b.backward_pointers == null) return (null); // Initialize the markers for backward pointers with INVALID_INDEX int backward_markers[] = new int[bn.number_variables()]; for (i = 0; i < backward_markers.length; i++) backward_markers[i] = BayesNet.INVALID_INDEX; // Initialize the marker for the last bucket backward_markers[b.variable.get_index()] = (int) (b.backward_pointers.get_value(0) + 0.5); // Go backwards through the bucket_tree for (i = (bi - 1); i >= 0; i--) { if (!bucket_tree[i].is_explanation()) break; back_df = bucket_tree[i].backward_pointers; // Skip null pointers (caused by evidence) if (back_df == null) continue; // Special treatment for bucket with only one value, // since it can be a bucket with only the bucket variable left if (back_df.number_values() == 1) { backward_markers[bucket_tree[i].variable.get_index()] = (int) (back_df.get_value(0) + 0.5); continue; } // Process the bucket j = back_df.get_position_from_indexes(bn.get_probability_variables(), backward_markers); backward_markers[bucket_tree[i].variable.get_index()] = (int) (back_df.get_value(j) + 0.5); } return (backward_markers); }
/* * Eliminates all variables defined as evidence. * The order of the variables that are not eliminated is * the same order in the original function. */ private ProbabilityFunction check_evidence(ProbabilityFunction pf) { int i, j, k, v, aux_i; boolean markers[] = new boolean[bn.number_variables()]; int n = build_evidence_markers(pf, markers); // Handle special cases if (n == 0) return (null); // No variable remains if (n == pf.number_variables()) return (pf); // No relevant evidence // Calculate necessary quantities in such a // way that the order of variables in the original // function is not altered. int joined_indexes[] = new int[n]; for (i = 0, j = 0, v = 1; i < pf.number_variables(); i++) { aux_i = pf.get_variable(i).get_index(); if (markers[aux_i] == true) { joined_indexes[j] = aux_i; j++; v *= bn.get_probability_variable(aux_i).number_values(); } } // Create new function to be filled with joined variables ProbabilityFunction new_pf = new ProbabilityFunction(bn, n, v, null); for (i = 0; i < n; i++) new_pf.set_variable(i, bn.get_probability_variable(joined_indexes[i])); // Loop through the values check_evidence_loop(new_pf, pf); return (new_pf); }
/* * Build an array of markers. The marker for a * variable is true only if the variable is present in the * input ProbabilityFunction pf and is not observed. * Even explanatory variables can be observed and taken as * evidence. */ private int build_evidence_markers(ProbabilityFunction pf, boolean markers[]) { int i, n; // Initialize the markers for (i = 0; i < markers.length; i++) markers[i] = false; // Insert the variables of the ProbabilityFunction for (i = 0; i < pf.number_variables(); i++) markers[pf.get_index(i)] = true; // Take the evidence out for (i = 0; i < bn.number_variables(); i++) { if (bn.get_probability_variable(i).is_observed()) markers[i] = false; } // Count how many variables remain n = 0; for (i = 0; i < markers.length; i++) { if (markers[i] == true) n++; } return (n); }
/* * Obtain the values for the evidence plus function. */ private void check_evidence_loop(ProbabilityFunction new_pf, ProbabilityFunction pf) { int i, j, k, l, m, p, last, current; int indexes[] = new int[bn.number_variables()]; int value_lengths[] = new int[bn.number_variables()]; for (i = 0; i < bn.number_variables(); i++) { indexes[i] = 0; value_lengths[i] = bn.get_probability_variable(i).number_values(); } for (i = 0; i < bn.number_variables(); i++) { if (bn.get_probability_variable(i).is_observed()) { indexes[i] = bn.get_probability_variable(i).get_observed_index(); } } last = new_pf.number_variables() - 1; for (i = 0; i < new_pf.number_values(); i++) { p = new_pf.get_position_from_indexes(indexes); new_pf.set_value(p, pf.evaluate(indexes)); indexes[new_pf.get_index(last)]++; for (j = last; j > 0; j--) { current = new_pf.get_index(j); if (indexes[current] >= value_lengths[current]) { indexes[current] = 0; indexes[new_pf.get_index(j - 1)]++; } else break; } } }
/** * Constructor for BucketTree. Does the whole initialization; it should be the only method that * deals with symbolic names for variables. */ public BucketTree(Ordering ord, boolean dpc) { int i, j, markers[]; ProbabilityFunction pf; ProbabilityVariable pv; DiscreteVariable aux_pv; DiscreteFunction ut; String order[]; do_produce_clusters = dpc; ordering = ord; // Collect information from the Ordering object. bn = ord.bn; explanation_status = ord.explanation_status; order = ord.order; // Indicate the first bucket to process active_bucket = 0; // Check the possibility that the query has an observed variable i = bn.index_of_variable(order[order.length - 1]); pv = bn.get_probability_variable(i); if (pv.is_observed() == true) { pf = transform_to_probability_function(bn, pv); bucket_tree = new Bucket[1]; bucket_tree[0] = new Bucket(this, pv, do_produce_clusters); insert(pf); } else { // Initialize the bucket objects bucket_tree = new Bucket[order.length]; for (i = 0; i < order.length; i++) { j = bn.index_of_variable(order[i]); bucket_tree[i] = new Bucket(this, bn.get_probability_variable(j), do_produce_clusters); } // Insert the probability functions into the bucket_tree; // first mark all functions that are actually going // into the bucket_tree. markers = new int[bn.number_variables()]; for (i = 0; i < order.length; i++) markers[bn.index_of_variable(order[i])] = 1; // Now insert functions that are marked and non-null. for (i = 0; i < bn.number_probability_functions(); i++) { if (markers[bn.get_probability_function(i).get_index(0)] == 1) { pf = check_evidence(bn.get_probability_function(i)); if (pf != null) { aux_pv = (bn.get_probability_function(i)).get_variable(0); insert(pf, !pf.memberOf(aux_pv.get_index())); } } } // Insert the utility_function. ut = bn.get_utility_function(); if (ut != null) insert(ut); } }
/** * Distribute evidence in the BucketTree. * * @return true if successful; false if not. */ public boolean distribute() { int i, j; boolean mark_non_conditioning[] = new boolean[bn.number_variables()]; // First make sure the BucketTree has been reduced. if (unnormalized_result == null) reduce(); // Second make sure there is more than one Bucket in the BucketTree. int last = bucket_tree.length - 1; if (last < 1) return (true); // Third, this method is used only if do_produce_clusters is true. if (do_produce_clusters == false) return (false); // Fourth, this method is use only if no explanatory variable was max'ed out. if (backward_pointers != null) return (false); // Go through the Bucket objects, from bottom to top, // to compute the new separator and cluster for each bucket. for (i = (last - 1); i >= 0; i--) { // Start from (last-1); last does not have child. // Check whether the Bucket has any valid content. if (bucket_tree[i].cluster == null) break; // Take the non-conditioning variables in a boolean array. for (j = 0; j < mark_non_conditioning.length; j++) mark_non_conditioning[j] = true; // OBS: The following piece of code will actually be less efficient than // necessary. It will count as "conditioning" any variable in the cluster // except the bucket variable. This will imply that some variables in the // separator will be normalized over without need, and the separator will // be larger than necessary. // OBS: this code was contributed by Wei Zhou ([email protected]), // who also detected the problem with the original code. // if (bucket_tree[i].cluster.number_variables() > // bucket_tree[i].non_conditioning_variables.size()) for (j = 1; j < bucket_tree[i].cluster.number_variables(); j++) { mark_non_conditioning[(bucket_tree[i].cluster.get_variables())[j].get_index()] = false; } // The following piece of code does the right thing (compared to the // piece of code above): it selects the // minimum number of non-conditioning variables. To use this piece // of code, it will be necessary to create a "normalize" method that // normalizes with respect to a number of variables at at time. /* for (j=0; j<bucket_tree[i].cluster.number_variables(); j++) { mark_non_conditioning[ (bucket_tree[i].cluster.get_variables())[j].get_index() ] = false; } for (Enumeration e = bucket_tree[i].non_conditioning_variables.elements(); e.hasMoreElements(); ) { ProbabilityVariable pv = (ProbabilityVariable)(e.nextElement()); mark_non_conditioning[pv.get_index() ] = true; } */ // Update the separator. bucket_tree[i].separator = bucket_tree[i].child.cluster.sum_out( bn.get_probability_variables(), mark_non_conditioning); // Compute cluster using new separator (note that if separator // is null, the cluster had all variables already processed). if (bucket_tree[i].separator != null) { // OBS: the method here should normalize with respect to more // than one variable, to allow this algorithm to be more efficient! bucket_tree[i].cluster.normalize_first(); // Now combine the cluster and the separator. bucket_tree[i].cluster = bucket_tree[i].cluster.multiply( bn.get_probability_variables(), bucket_tree[i].separator); } // Mark the Bucket as DISTRIBUTED. bucket_tree[i].bucket_status = Bucket.DISTRIBUTED; } // Indicate success. return (true); }