Here is a multi-criteria approach to identify and evaluate network-flow monitoring strategies in a stochastic environment. We propose an analytical approach for assessing flow disturbance, based on limited sampling of arc-flow information in multi commodity, or multiple origin-destination (O-D), networks. A disturbance is defined here as abnormal network traffic beyond routine fluctuations. The network is characterized by arcs that are available with certain probabilities, suggesting a busy packet-switching/circuit-switching network, or simply network with failing components. Given the huge state space of such stochastic networks, our first objective is to bound the expected flow for computational efficiency. This is accomplished by extending available single-commodity-network results to the multi commodity case. The second objective is to determine the best placement of flow monitors to obtain the most accurate estimates of traffic flow, as represented by O-D pair traffic volumes in order to detect a disturbance. We followed a multi-criteria approach in A. R. Robinson defining and evaluating all possible monitor-placement strategies satisfying monitor availability. To assess a traffic pattern, O-D volumes are estimated using the l p -metric for p = 1, 2 and ∞, representing a decisionmaker's risk-preference structure. The final objective is to define a disturbance metric providing confident assessments on the occurrence of a flow disturbance. In this regard, we use linear-regression to generate confidence intervals around the expected flow for each O-D pair. This is supplemented by the Fisher Method of Randomization, a non-parametric test, for small networks. Unlike the Fisher-test results, it appears the linear-regression "disturbance metric" is very sensitive to detecting "disturbance." Preliminary results suggest that among the three l p metrics in O-D estimation, it appears that the best is the l 2 -metric, which provides the appropriate sensitivity and accuracy. Counter to intuition, it is also observed that the "disturbance" estimation error generally decreases as the number of sampled arcs decreases.