Estimating quality of service (QoS) parameters such as link delay distribution from the end-to-end delay of a multicast tree topology in network tomography cannot be achieved without multicast probing techniques or designing unicast probing packets that mimic the characteristics of multicast probing packets. Active probing is gradually giving way to passive measurement techniques. With the emergence of next generation networks such as Autonomic Network Architecture (ANA) network, which do not support active probing, a new way of thinking is required to provide network tomography support for such networks. This thesis is about investigating the possible solution to such problem in network tomography. Two approaches, queue model and adaptive learning model were implemented to minimize the uncertainty in the end-to-end delay measurements from passive data source so that we could obtain end-to-end delay measurements that exhibit the characteristics of unicast or multicast probing packets. The result shows that the adaptive learning model performs better than the queue model. In spite of its good performance against the queue model, it fails to outperform the unicast model. Overall, the correlation between the adaptive learning model and multicast probing model is quite weak when the traffic intensity is low and strong when the traffic intensity is high. The adaptive model may be susceptible to low traffic. In general, this thesis is a paradigm shift from the investigation of "deconvolution" algorithms that uncover link delay distributions to how to estimate link delay distributions without active probing.