Method replay

tf2.0/mlp_trader.py:269–300

(self, batch_size=32)

    return np.argmax(act_values[0])  # returns action

  def replay(self, batch_size=32):
    # first check if replay buffer contains enough data
    if self.memory.size < batch_size:
      return

    # sample a batch of data from the replay memory
    minibatch = self.memory.sample_batch(batch_size)
    states = minibatch['s']
    actions = minibatch['a']
    rewards = minibatch['r']
    next_states = minibatch['s2']
    done = minibatch['d']

    # Calculate the tentative target: Q(s',a)
    target = rewards + (1 - done) * self.gamma * np.amax(self.model.predict(next_states), axis=1)

    # With the Keras API, the target (usually) must have the same
    # shape as the predictions.
    # However, we only need to update the network for the actions
    # which were actually taken.
    # We can accomplish this by setting the target to be equal to
    # the prediction for all values.
    # Then, only change the targets for the actions taken.
    # Q(s,a)
    target_full = self.model.predict(states)
    target_full[np.arange(batch_size), actions] = target

    # Run one training step
    self.model.partial_fit(states, target_full)

    if self.epsilon > self.epsilon_min:
      self.epsilon *= self.epsilon_decay


  def load(self, name):