bugfix: don't sample action distribution twice

there was a nasty bug that caused the controller to sample
from the softmax action probability distribution twice.
this is problematic because the action choice determines
both the log probability of the action as well as the
choice of the next action_classifier

deep_cash/deep_cash/cash_controller.py

@@ -143,6 +143,8 @@ def decode(
 
         Where the actions are a sequence of algorithm components and
         hyperparameter settings.
+
+        TODO: add unit tests for this method and related methods.
         """
         input_tensor, hidden = init_input_tensor, init_hidden
         actions = []
@@ -173,7 +175,7 @@ def select_action(self, action_probs, action_index):
         action_classifier = self.action_classifiers[action_index]
         action_dist = Categorical(action_probs)
         choice_index = action_dist.sample()
-        _choice_index = int(action_dist.sample().data)
+        _choice_index = int(choice_index.data)
         return {
             "action_type": action_classifier["action_type"],
             "action_name": action_classifier["name"],