Etude et application de systèmes hybrides neurosymboliques
Bruno Orsier · HAL (Le Centre pour la Communication Scientifique Directe) · 1995
For several years, hybrid neurosymbolic systems (HNSS) have combined artificial neural networks with symbolic systems (knowledge-based systems), in order to take advantage of their strengths. We present a taxonomy of the numerous existing HNSS, according to two perspectives: the degree of coupling (weak, tight, strong), and the interaction type (pre/postprocessing, subprocessing, coprocessing, metaprocessing). The taxonomy also embodies two other less hybrid approaches of neurosymbolic integration: one is purely neural and the other is translation-based. Then we study an existing HNSS, SYNHESYS, and we examine its application possibilities in micro-electronics and alpine geography. Then we propose a new hybrid architecture, NESSY3L, and its application to the reactive control of a mobile robot. The architecture consists of three levels, purely neural, neurosymbolic, purely symbolic, and it has been integrated into the robot simulator MOLUSC. This architecture offers interesting prospects for the HNSS, among which are i) taking time into account, ii) development of inter-levels mechanisms and evolution towards strong coupling, iii) use of ideas from other approaches of neurosymbolic integration.