cartpole_double_dqn_experience_replay.py

def experience_replay(self):

    minibatch = random.sample(self.memory, EXPERIENCE_REPLAY_BATCH_SIZE)
    minibatch_new_q_values = []

    for experience in minibatch:
        state, action, reward, next_state, done = experience
        state = self._reshape_state_for_net(state)
        experience_new_q_values = self.online_network.predict(state)[0]
        if done:
            q_update = reward
        else:
            next_state = self._reshape_state_for_net(next_state)
            # using online network to SELECT action
            online_net_selected_action = np.argmax(self.online_network.predict(next_state))
            # using target network to EVALUATE action
            target_net_evaluated_q_value = self.target_network\
                .predict(next_state)[0][online_net_selected_action]
            q_update = reward + GAMMA * target_net_evaluated_q_value
        experience_new_q_values[action] = q_update
        minibatch_new_q_values.append(experience_new_q_values)
    minibatch_states = np.array([e[0] for e in minibatch])
    minibatch_new_q_values = np.array(minibatch_new_q_values)
    self.online_network.fit(minibatch_states, minibatch_new_q_values, verbose=False, epochs=1)