Corrected bug where system wanted to back at each end of sentence

Train.py

@@ -254,9 +254,9 @@ def trainModelRl(debug, networkName, modelDir, filename, nbIter, batchSize, devF
         reward = torch.FloatTensor([reward_]).to(getDevice())
 
         newState = None
+        toState = strategy[action.name][1] if action.name in strategy else -1
         if appliable :
           applyTransition(strategy, sentence, action, reward_)
-          toState = sentence.state
           newState = policy_net.extractFeatures(dicts, sentence).to(getDevice())
         else:
           count+=1