Understanding and predicting of human driving behavior play an important role in the development of Advanced Driver Assistance Systems (ADAS) for assisting drivers. In this contribution, a Multi-Layer (3-layer) Hidden Markov Models (HMM) approach is proposed and developed for predicting human driving behavior. For a single HMM, more inputs will cause a longer training time, higher complexity, and even overfitting. The proposed method can fit to complex situations, also when more inputs are considered. In this contribution the first layer is considered to predict driving behavior in certain single working cases, i.e. each input variable is used to train a single model independently in the first layer. The outputs are combined into different models containing different information in the second and third layers. All HMMs in combination with a prefilter are used to predict driving behavior in parallel. Lane changing, as a usual driving maneuver, will be used as representative example task to be predicted. Based on observations (training), the HMM algorithm calculates the most probable driving behavior through the observation sequences. Furthermore, the observed sequences are also used for training of HMM during modeling process.To define model parameters and to improve the model performance NSGA-II is used. Using experimental data taken from driving simulator, it can be concluded that selecting optimal parameters increase the performance of driving behavior prediction. The effectiveness of the suggested Multi-Layer HMMs has been successfully proved based on experiments. The results show that the newly introduced approach outperforms alternative approaches applied to the same data set.