EPSRC logo

Details of Grant 

EPSRC Reference: GR/H85427/01
Title: NETWORK PROGRAMME FOR SOFTWARE RELIABILITY -IED4/1/9301
Principal Investigator: Partridge, Professor D
Other Investigators:
Sharkey, Professor N
Researcher Co-Investigators:
Project Partners:
Department: Computer Science
Organisation: University of Exeter
Scheme: Standard Research (Pre-FEC)
Starts: 01 February 1993 Ends: 31 July 1996 Value (£): 401,107
EPSRC Research Topic Classifications:
Software Engineering
EPSRC Industrial Sector Classifications:
Information Technologies
Related Grants:
Panel History:  
Summary on Grant Application Form
1. To quantify the diversity, and hence system reliability, achievable by including trained neural-net versions in a conventional multiversion system.2. To measure the diversity achievable between neural-net versions.3. To develop a formal basis for 'network programming'.4. To assess the reliability of network programmed systems.Progress:The Exeter group is making steady progress in implementing non-trivial functions as reliable neural-net systems. We have developed (with considerable help from Professor Krzanowski) a practical statistical basis to both assist in the engineering (e.g. by comparing intra- and inter-set diversity measures) and validation of complex multiversion systems i.e. composed of sets of version sets. We have made considerable progress with various of the elements of the novel methodology (see below) - in particular with the specification of (and thus differentiation between) optimal training, acceptance and validation sets. We are also making good progress with two-level structures (i.e. sets of version sets) for complex multiversion systems design - and the result is systems that are substantially more reliable than any of their components and are extremely robust. A variety of results have been written up and published. Several more papers are currently under review in neural net journals.A radically new software development paradigm is emerging - multiversion neural net systems. It is based on exploitation of the fact that: neural nets are quickly and cheaply trained; function implemented is determined by initial conditions of training; data-defined problems can be implemented prior to (abstract) specification development; implementation is a well-defined process (algorithmic) based on well-defined data. It avoids the known weaknesses of neural computing: it is an 'approximating' technology; single implementations are effectively non-replicable; operational details of non-trivial problems are incomprehensible.In addition to the conventional minimum coincident failure diversity, we are exploring several different types of diversity. Of particular interest is the diversity introduced by noise , such as randomisation of initial link weights, which can be exploited by averaging version outcomes. This possibility, which appears to be most productive, is dependent upon the approximating nature of neural computing. Thus it appears that certain strategies for reliability improvement are peculiar to neural computing and do not exist in conventional software technology.
Key Findings
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Potential use in non-academic contexts
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Impacts
Description This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Summary
Date Materialised
Sectors submitted by the Researcher
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Project URL:  
Further Information:  
Organisation Website: http://www.ex.ac.uk