The INFOCOM'08 paper titled "Minerva: Learning to Infer Network Path Properties" written by Wouhaybi et al. proposes to predict network path properties based on learning techniques.

This last decade, plenty of solutions have been proposed to estimate the path properties between nodes without having to explicitly measure the path between the two nodes. However, most of the techniques (with or without landmarks) require to periodically make some measurements.

The authors propose to use well-known learning techniques to infer network properties between nodes. They propose to use Bayesian network to learn and infer the properties. The idea is simple, just use some landmarks and measure the path from the nodes to the landmarks. Then divide the measurements in histograms. The Bayesian network learn the probabilities of the path properties and can return the probability for each value in the histogram of any node pair. The value with the highest probability corresponds to the most probable value of the path property.

The evaluation shows that Minerva is better than Vivaldi or Min-Sum. For instance, the latency error is less than 10ms for 80% of the cases in their experiments (don't know the real experiments however...) while it is 160ms for Vivaldi.

The problem with Minerva is the need of learning and re-learning. In the paper it is stated that it is not really necessary to make re-learning but the evaluation seems to be limited. In addition no real clue is given to estimate the cost of the learning phase.

Despite the last remarks, Minerva seems to be a promising technique and sure that we'll see improvements of that technique in the future.

Paper available at: http://www.cs.dartmouth.edu/~campbell/minvera.pdf

This paper is related to our IDIPS researches (http://inl.info.ucl.ac.be/idips).