Learning Abstract Behaviors with the Hierarchical Incremental Gaussian Mixture Network

This paper presents a new probabilistic hierarchical model, called HIGMN (Hierarchical Incremental Gaussian Mixture Network), which is based on ideas presented by Deep Architectures. The proposed model, composed by layers of IGMNs, is able to extract features from data input of different domains in the low-level layers and to correlate these features in a high-level layer. Experiments show that HIGMN is able to learn an abstract behavior using the features extracted from sensory and motor data of a mobile robot and to perform correct actions even in unknown instances of sensory perception.