generator.py 文件源码

def _build_discounts_matrix(self, T, gamma):
        """Build lower-triangular matrix of discounts.
        For example for T = 3: D = [[1,       0,     0]
                                   [gamma,   1,     0]
                                   [gamma^2, gamma, 1]]
        Then with R, our N x T incremental rewards matrix, the discounted sum is
            R * D
        """
        power_ltri  = tf.cumsum(
            tf.sequence_mask(tf.range(T)+1, T, dtype=tf.float32), exclusive=True
        )
        gamma_ltri  = tf.pow(gamma, power_ltri) 
        gamma_ltri *= tf.sequence_mask(tf.range(T)+1, T, dtype=tf.float32)
        return gamma_ltri