Hierarchical concept learning based on functional similarity of actions

Imitation is different from other similar social learning techniques such as mimicking. True imitation is goal-directed and involves abstraction. This paper presents a hierarchical concept learning scheme, which uses imitation to learn new actions based on perceived observations. Presented demonstrations are composed of a spatio-temporal signal and a feature vector. In this paper a probabilistic model is provided in order to represent these hybrid demonstrations. These models have been used in an autonomous concept learning algorithm which uses effects of actions as a measure of functional similarity. Also a functional space is defined to determine concepts similarity. An algorithm for high-level concept formation in this space is provided. The proposed methods are used and evaluated in an air hockey simulation environment. Simulation results show efficiency of the proposed methods in recognition and reproduction of air hockey basic shots and their ability for hierarchical concept learning.