The goal of this project is to advance the state of the art in symbolic-connectionist integration (SCI), and the work will aim to:
- enhance the versatility of symbolic-connectionist hybrid systems (SCSs) by combining a variety of representation, inferencing and learning schemes from the symbolic and the connectionist paradigms
- augment the reasoning power of SCSs by overcoming their traditional limitation to propositional logic
- clarify the theoretical impact of symbolic-connectionist integration on fundamental machine learning issues such as the choice and combination of learning methods, credit assignment and knowledge assimilation
- ensure the scalability and real-world applicability of the resulting hybrid models by testing them on prediction, optimisation and classification problems in industrial and other areas.
APPROACH AND METHODS
Current SCSs are small experimental systems which ally one symbolic and one connectionist model using simple, often ad hoc coupling modes and techniques. To blend a wider range of models from the two paradigms within a coherent system, a principled approach is required which takes account of final integration needs in the design of the individual components. We have chosen a distributed approach to SCI. An initial phase of the project will consist in the specification and implementation of a distributed architecture for the cooperation of multiple heterogeneous agents. At the outset generic agents will draw from any of the processing models within the two paradigms to accomplish their given tasks. With problem-solving experience, however, each will specialise more and more on methods best adapted to its specific problem context. These methods will be based on a broad repertoire of inferencing and learning strategies which will be made available by the hybrid models built in the project. These hybrid models will be created in a modular and incremental fashion. Neural networks will be combined with fuzzy logic, case-based reasoning to form partial hybrids, which will then be integrated within a single unified model.
A result of this project will be insight into the complex problems of SCI. In the field of machine learning, where multistrategy learning remains essentially symbolic, this will be the first medium-scale effort we know of to implement multiparadigm, multistrategy learning in knowledge-based systems. From a practical point of view, hybrid models should give a new impetus to the incorporation of AI techniques in industrial applications where purely symbolic or purely connectionist processing models have been found wanting. Advances in SCI are also expected to have an impact on the software engineering industry.
1211 Geneva 4