This study is an attempt to design a method for an autonomous pattern classification and recognition system for emotion recognition. The proposed system utilizes Biased ARTMAP for pattern learning and classification. The ARTMAP system is dependent on training sequence presentation to determine the effectiveness of the learning processes, as well as the strength of the biasing parameter, lambda
. The optimal combination of
and training sequence can be computed efficiently using a genetic permutation algorithm. The best combinations were selected to train individual ARTMAPs as voting members, and the final class predictions were determined using probabilistic ensemble voting strategy. Classification performance can be improved by implementing a reliability threshold for training data. Reliability metric for each training sample was computed from the current voter output, and unreliable training samples were excluded from the performance calculation. Individual emotional states are highly variable and are subject to evolution from personal experiences. For this reason, the above system is designed to be able to perform learning and classification in real-time to account for inter-individual and intra-individual emotional drift over time.