Towards Developing Verifiable Neural Network Controller

Artificial Neural Networks(ANN) play an important part in developing intelligent and autonomous systems. By training the ANN with desired input-output patterns that are derived from human experience or simulation, the designers of such system avoid specifying detailed analytical model of such complex systems. Unlike conventional system design and development which is based on an iterative procedure of system analysis, specification, implementation and testing, ANN based system relies on training to formulate the control mechanisms. When such ANN-based components are embedded in a larger system, their interactions become harder to analyze and model. Formal testing of such system for safety properties is extremely hard due to the lack of a complete system model. In this paper, we propose the neuralware engineering framework to address the above issues. It is based on an iterative approach on specification, model checking, and testing. After the ANN-based system is designed and trained using an initial partial system model, an rule extraction algorithm is used to discover what has been learned. The discrepancies between the learned rules and the model is compared to modify the system model. We present in this paper a simple example as a proof of concept of our framework.

[1]  Mandayam K. Srivas,et al.  A Tutorial Introduction to PVS , 1998 .

[2]  Chuen-Tsai Sun,et al.  Neuro-fuzzy modeling and control , 1995, Proc. IEEE.

[3]  Bart Kosko,et al.  Adaptive fuzzy systems for backing up a truck-and-trailer , 1992, IEEE Trans. Neural Networks.

[4]  E. R. Khan.  NeuFuz: an intelligent combination of fuzzy logic with neural nets , 1993, Proceedings of 1993 International Conference on Neural Networks (IJCNN-93-Nagoya, Japan).

[5]  Stephen J. Andriole,et al.  Software validation, verification, testing, and documentation , 1986 .

[6]  Stephen J. Andriole.  Software Validation, Verification, Testing and Documentation: A Source Book , 1986 .

[7]  George Cybenko,et al.  Approximation by superpositions of a sigmoidal function , 1989, Math. Control. Signals Syst..

[8]  Constance L. Heitmeyer,et al.  Consistency checking of SCR-style requirements specifications , 1995, Proceedings of 1995 IEEE International Symposium on Requirements Engineering (RE'95).

[9]  Michael Kincheloe,et al.  On-line learning neural-network controllers for autopilot systems , 1995 .

[10]  Huan Liu,et al.  Symbolic Representation of Neural Networks , 1996, Computer.

[11]  John R. Callahan,et al.  An approach to verification and validation of a reliable multicasting protocol , 1996, ISSTA '96.

[12]  Nadipuram R. Prasad,et al.  Neural Networks and Fuzzy Logic , 1998 .

[13]  Steve Easterbrook,et al.  Scr as an Iv&v Tool , 1996 .

[14]  Kurt Hornik,et al.  Approximation capabilities of multilayer feedforward networks , 1991, Neural Networks.

[15]  Standard Glossary of Software Engineering Terminology , 1990 .

[16]  John C. Cherniavsky,et al.  Validation, Verification, and Testing of Computer Software , 1982, CSUR.

[17]  Wu Wen,et al.  Neuralware engineering: develop verifiable ANN-based systems , 1996, Proceedings IEEE International Joint Symposia on Intelligence and Systems.