On Machine Symbol Grounding and Optimization

Abstract

From the point of view of an autonomous agent the world consists of high-dimensional dynamic sensorimotor data. Interface algorithms translate this data into symbols that are easier to handle for cognitive processes. Symbol grounding is about whether these systems can, based on this data, construct symbols that serve as a vehicle for higher symbol-oriented cognitive processes. Machine learning and data mining techniques are geared towards finding structures and input-output relations in this data by providing appropriate interface algorithms that translate raw data into symbols. This work formulates the interface design as global optimization problem with the objective to maximize the success of the overlying symbolic algorithm. For its implementation various known algorithms from data mining and machine learning turn out to be adequate methods that do not only exploit the intrinsic structure of the subsymbolic data, but that also allow to flexibly adapt to the objectives of the symbolic process. Furthermore, this work discusses the optimization formulation as a functional perspective on symbol grounding that does not hurt the zero semantical commitment condition. A case study illustrates technical details of the machine symbol grounding approach.