Unsupervised test model reconstruction from conformance test logs

The high manpower cost of test model generation is a big obstacle for model based testing. In this paper, we propose a method that decreases this cost for the case of performance testing of telecommunication systems. The test model which is typically a deterministic finite state machine is partially derived from execution logs of black box system tests by means of statistical methods and then manually minimized. A sequential pattern mining algorithm is used for identifying frequently observed event sequences, from which it is possible to identify partitions of states of the machine and set up functional dependencies among them. Our method in contrast to other methods is passive and unsupervised it does neither require access to the black box system to be learnt, nor the input sequences to be finite. However, the equivalence of some states cannot be proven, therefore the model has to be finalized manually. We show experimental results on random incompletely specified FSMs and sets of random walk traces to evaluate the efficiency of our method.