fix(a3c): remove root noise from action selector

 - the idea was to include noise at the beginning of episodes, sort of how Zero friends do it.
 - instead just use eGreedy like the selector says

Author: justindujardin

libraries/mathy_python/mathy/agents/action_selectors.py

@@ -48,9 +48,6 @@ class A3CEpsilonGreedyActionSelector(ActionSelector):
     def __init__(self, *, epsilon: float, **kwargs):
         super(A3CEpsilonGreedyActionSelector, self).__init__(**kwargs)
         self.epsilon = epsilon
-        self.noise_alpha = 0.3
-        self.use_noise = True
-        self.first_step = True
 
     def select(
         self,
@@ -63,16 +60,6 @@ def select(
 
         probs, value = self.model.predict_next(last_window.to_inputs())
         last_move_mask = last_window.mask[-1]
-        # Apply noise to the root node (like AlphaGoZero MCTS)
-        if self.use_noise is True and self.first_step is True:
-            noise = np.random.dirichlet([self.noise_alpha] * len(probs))
-            noise = noise * np.array(last_move_mask)
-            probs += noise
-            pi_sum = np.sum(probs)
-            if pi_sum > 0:
-                probs /= pi_sum
-            self.first_step = False
-
         no_random = bool(self.worker_id == 0)
         if not no_random and np.random.random() < self.epsilon:
             # Select a random action