Skip to main content

Integration of Symbolic and Numeric Learning Techniques

Objective

The objective of INSTIL was to improve knowledge acquisition for knowledge-based systems by the application of machine learning techniques. The method of knowledge acquisition to be developed was the generation of a knowledge-base from an analysis of examples of sets of field data by the formulation of rules in the application domain.
The project set out to identify the best features from different approaches to knowledge acquisition, in order to improve the quality of rules extracted from noisy data. Use was made of existing software: MAIN, based on Michalski's INDUCE and AQ11, a symbolic learner; AGAPE and AGAPE-C, which use theorem-proving techniques and taxonomies of descriptors of example sets; and NEDDIE, an extended version of Quinlan's ID3, which uses numerical manipulation to constrain its search space.
The objective was to improve knowledge acquisition for knowledge-based systems by the application of machine learning techniques. The method of knowledge acquisition developed was the generation of a knowledge base from an analysis of examples of sets of field data by the formulation of rules in the application domain. The project set out to identify the best features from different approaches to knowledge acquisition, in order to improve the quality of rules extracted from noisy data. Use was made of existing software: MAIN, a symbolic learner; AGAPE and AGAPE-C, which use theorem providing techniques and taxonomies of descriptors of example sets; and NEDDIE, which uses numerical manipulation to constrain its search space. Improved versions were produced of the programs and ported to the chosen development environment. Prototypes integrating MAIN, NEDDIE and AGAPE were completed and distributed for experimentation. The most promising prototype used MAGGY (augmented AGAPE) to generalize cluster descriptions obtained by NEDDIE. This prototype forms the basis of the system, which is now being finalized and documented. An object oriented representation language, the generalization oriented language (GOL), was implemented and documented. It is used to represent background knowledge, examples, and the rules synthesised by the system. The integrated learning system is being strengthened by studies on dealing with noise in knowledge acquisition. The system has ben tested on several large-scale applications in various domains including agriculture (diagnosis of disease in crops), image understanding (object recognition, mosquito recognition, medical diagnosis, air traffic control, and fault recognition in turbine generators.
An example application was demonstrated to show integrated learning of rules using symbolic and numerical methods.
Improved versions were produced of the programs previously developed separately by project team members at their locations, and ported to the chosen development environment. Prototypes integrating MAIN, NEDDIE and AGAPE were completed and distributed to the partners for experimentation. The most promising prototype used MAGGY (augmented AGAPE) to generalise cluster descriptions obtained by NEDDIE. This prototype forms the basis of the INSTIL system, which is now being finalised and documented. An object-o riented representation language, the generalisation-oriented language (GOL), was implemented and documented. It is used to represent background knowledge, examples, and the rules synthesised by INSTIL. The integrated learning system is being strengthened by studies on dealing with noise in knowledge acquisition.
The INSTIL system has been tested on several large-scale applications in various domains including agriculture (diagnosis of disease in crops), image understanding (object recognition), mosquito recognition, medical diagnosis, air traffic control, and fault recognition in turbine generators.
Exploitation
It is expected that the lessons learned will be incorporated in tool environments for expert system construction in the form of an automatic rule refinement and acquisition module. The final prototype of this module and a complete user manual is now available.
Cognitech intends to enhance the set of expert systems under development for diagnosing disease in 30 different crops by the addition of an automatic rule-refinement and acquisition module. Cognitech also intends to incorporate a similar rule-learning module in their integrated laboratory for teaching AI.
GEC-RL intends to link a rule-learning element from INSTIL to a tool environment for expert system construction with which they are involved.
Test sites of an industrial, commercial, or medical nature are being sought for the INSTIL system. GEC and the University of Paris are using INSTIL in traffic control, case law deduction, and the identification of address locations on parcels.
Results from INSTIL have been incorporated in ESPRIT project 2154, MLT (Machine Learning Toolbox).

Coordinator

GEC Marconi Research Centre
Address
West Hanningfield Road Great Baddow
CM2 8HN Chelmsford
United Kingdom

Participants (2)

Framentec Cognitech SA
France
Address
Tour Fiat 1 Place De La Coupole
92084 Paris La Défense
Université de Paris XI (Université Paris-Sud)
France
Address
Avenue Georges Clémenceau
91405 Orsay