In this paper we present a work on encoding, clustering and modeling expert driver behaviors to be reproduced by an autonomous driving car agent. We collected speed, brake, steering wheel and acceleration data from an expert driving in typical Japanese suburban areas. We grouped together consecutive data in a sliding window as the input to a fully connected deep autoencoder for dimension reduction. We then clustered the encoded driving data into 9 different behaviors. For each behavior class, we modeled the speed using a fit of same clustered data and tested the models in a ROS car simulator. Results show that the models created from the clusters could represent different driving behaviors on various roads. Experimental simulations show that the proposed approach can reproduce speed behavior comparable to that of the expert drivers even in new environments.