75eeb9825e331581839ad647acd15a40cd250c35,tensorforce/value_functions/deep_q_network.py,DeepQNetwork,create_training_operations,#DeepQNetwork#,95

Before Change

        with tf.name_scope("predict"):
            self.dqn_action = tf.argmax(self.training_network(self.state), dimension=1, name="dqn_action")

        with tf.name_scope("target_values"):
            self.target_values = tf.reduce_max(self.target_network(self.next_states), reduction_indices=1, name="target_values")

        with tf.name_scope("training"):
            self.q_targets = tf.placeholder("float32", [None], name="batch_q_targets")
            self.batch_actions = tf.placeholder("int64", [None], name="batch_actions")

After Change

            self.batch_next_states = tf.placeholder(tf.float32, None, name="next_states")
            self.batch_actions = tf.placeholder(tf.int64, [None], name="batch_actions")
            self.batch_terminals = tf.placeholder(tf.float32, [None], name="batch_terminals")
            self.batch_rewards = tf.placeholder(tf.float32, [None], name="batch_rewards")

            float_terminals = np.array(self.batch_terminals, dtype=float)
            self.target_values = tf.reduce_max(self.target_network(self.batch_next_states), reduction_indices=1,
                                               name="target_values")

            q_targets = self.batch_rewards + (1. - float_terminals) * self.gamma * self.target_values

            actions_one_hot = tf.one_hot(self.batch_actions, self.actions, 1.0, 0.0)

            self.batch_q_values = tf.identity(self.training_network(self.batch_states), name="batch_q_values")

Instances