Hierarchical Learning of Generative Automaton Models from Sequential Data

Abstract

Passive automata learning is a method for inferring automaton models from a given collection of observations of system behavior (traces). It has been applied to reactive systems with probabilistic behavior. In particular, IOAlergia is a well known algorithm for inferring models in the form of deterministically labeled Markov decision processes from system traces. The quality of the resulting model depends heavily on the provided data set and suffers if data is scarce. However, in many cases additional knowledge about the system is available. This work aims to incorporate knowledge about system modes into the learning process in order to improve model quality for low-data scenarios. This is done by splitting the traces according to system modes, learning individual models for each mode and combining those models into one model with sub-regions corresponding to individual system modes. In our evaluation on artificial models, our method outperforms the baseline in at least 90 % of cases for all considered metrics. This method was developed to learn generative models of human driving behavior. Data from recorded test drives on highways was used to learn a hierarchical stochastic model of typical acceleration behavior of human drivers. In the automotive industry, such models make the simulations of driving emissions more realistic.