Integration of Sub-Symbolic and Symbolic Information Processing in Robot Control

Abstract

In the Autonomous m b i l e Systems project, the FAW uses a mobile robot to study questions related t o the d e e p integration of sub-symbolic and symbolic information processing. AMOS aims at methods for autonomously acquiring new concepts via induction from the environment. AMOS is (deliberately) equipped with an incomplete model of itself and of the environment. The robot plans its actions in order t o perform certain $asks, e.g. visitiny certain locations. The successful execution of a plan results in positive reinforcement. When AMOS recognizes substantial differences between expectation and observation, it will collect and classify the available sensor information. FAW uses its sub-symbolic image processing system ALIAS to event,ually translate collections of such information into a new concept. Such a concept is then integrated into the symbolic world model of AMOS to improve the robot’s performance, while at the same time providing feedback concerning the appropriateness of concepts learned. 1 Basic Ideas behind the Project The conceptual outline of the project is motivated by aspects of the evolution of life on earth. In the course of evolution, sub-symbolic forms of information processing via neural networks have been of central importance. Essential steps have been the creation of mechanisms which are able to process sensor information (such as pixel images or other data streams) as a basis for behavior control. These steps can, to some extent, be interpreted as early forms of implicit concept generation. Based on collective learnirg system theory, the FAW projects ALIAS and ALA” have demonstrated, supplementary to other connectionistic approaches in this field, the ability to generate concepts carrying semantics in a static environment using a simple organizing principle: spatial neighborhood in images. This reflects one of the laws of ”Gestalt” which was long ago discovered by psychologist,s. The basis for an intelligent behavior of systems has gradually improved over the course of evolution. as the level of a mere processing of stimulus-response patterns was surpassed, and more abstract principles of identifying, organizing, and processing of such patterns emerged [all. Usually, one tries to capture and describe this more abstract level by notions such as classes, categorzes, or the notion of symbol and respective forms of information processing (e.g. , logical inferences). The gradual transition to ever broader forms of symbol processing can be seen as the decisive step in very compact forms of information coding and processing, which are, nevertheless, biologically realizable within a neural network (corresponding to the observation that most types of artificial neural networks allow, among other things, the emulation of (finite) Turing machines, cf. also [23] [24]). In spite of this importance of symbol processing, even today, the quite rare process of generating genuine new concepts (which is considered one of the most sophisticated abilities of particularly creative humans) often seems to be based more on sub-symbolic forms of mforniat ion processing (intuition, holistic understanding) tlian on symbolic forms, where both processing modes are often closely coupled. Given this observation, one of the most challenging aims of the AMOS project is to better understand this essential bootstrap phenomena of a deep integration of sub-symbolic and symbolic information processing. For present-day research in AI, the close interaction of sub-symbolic and symbolic forms of informa238 0-8186-2675-5/92 $3.00 Q 1992 IEEE tion processing is, consequently, a topic of basic importance. The approach discussed here does not aim to solve this integration question in the general sense. On the other hand, it also does not aim a t rather straightforward integration steps, such as merely representing symbolic concepts via neural networks, although this is a.n important basic aspect. l.nstead, the symbolic level in AMOS will manage contextual information for the sub-symbolic learning of completely new concepts in the sense that it allows a certain pre-classification of situations, and otherwise provides for the basic day-to-day operation of the system. This last aspect may be understood in the sense of controlling the regular behavior of the system via symbolic forms of information processing (e.g. planning, control) and, in addition, symbolic forms of learning (based on statistical methods). Given this basic set-up, the aim is then to show that another simple principle can adequately direct the process of collecting and merging data streams to create new concepts sub-symbolically; this principle is the sudden significant divergence of expectation and observation. Thus, the main organizing principle to create genuine new concepts will be the regularities supplied by real-life phenomena [3] [4]. Feedback with reality is so complex that whatever we would design as a model of the environment will not be an appropriate representation in the view of our robot experiment, i.e. simulation just will not do. In other words, to identify sub-symbolic concepts, deliberately hidden in a simulation set-up, is not what we would regard as appropriate. 2 The FAW Robot System