[Frederik Eaton] Improvements of Gibbs class

* [Frederik Eaton] Added Gibbs::state() accessors/mutators
* [Frederik Eaton] Fixed Gibbs::getVarDist(size_t) to return uniform
  distribution if no state is allowed

diff --git a/ChangeLog b/ChangeLog
index c1f8a37..e2175c3 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,6 @@
+* [Frederik Eaton] Added Gibbs::state() accessors/mutators
+* [Frederik Eaton] Fixed Gibbs::getVarDist(size_t) to return uniform 
+  distribution if no state is allowed
 * [Frederik Eaton] Improved parsing code in tests/testdai to allow recursive
   alias substitutions
 * Interface changes:

diff --git a/include/dai/gibbs.h b/include/dai/gibbs.h
index afdaa06..2e5ec68 100644 (file)
--- a/include/dai/gibbs.h
+++ b/include/dai/gibbs.h
@@ -88,13 +88,19 @@ class Gibbs : public DAIAlgFG {
         //@}
 
 
-        /// @name Additional interface specific for BP
+        /// @name Additional interface specific for Gibbs
         //@{
         Factor beliefV( size_t i ) const;
         Factor beliefF( size_t I ) const;
         void randomizeState();
         //@}
 
+        /// Return reference to current state vector
+        std::vector<size_t>& state() { return _state; }
+
+        /// Return const reference to current state vector
+        const std::vector<size_t>& state() const { return _state; }
+
     private:
         void updateCounts();
         Prob getVarDist( size_t i );

diff --git a/src/gibbs.cpp b/src/gibbs.cpp
index bbe3f0e..df55257 100644 (file)
--- a/src/gibbs.cpp
+++ b/src/gibbs.cpp
@@ -126,7 +126,7 @@ Prob Gibbs::getVarDist( size_t i ) {
     size_t i_states = var(i).states();
     Prob i_given_MB( i_states, 1.0 );
 
-    // use markov blanket of var(i) to calculate distribution
+    // use Markov blanket of var(i) to calculate distribution
     foreach( const Neighbor &I, nbV(i) ) {
         const Factor &f_I = factor(I);
         size_t I_skip = getFactorEntryDiff( I, i );
@@ -137,7 +137,13 @@ Prob Gibbs::getVarDist( size_t i ) {
         }
     }
 
-    return i_given_MB.normalized();
+    if( i_given_MB.sum() == 0.0 )
+        // If no state of i is allowed, use uniform distribution
+        // FIXME is that indeed the right thing to do?
+        i_given_MB = Prob( i_states );
+    else
+        i_given_MB.normalize();
+    return i_given_MB;
 }