| EPSRC Reference: |
GR/J14301/01 |
| Title: |
NOVEL PARALLEL GENETIC ALGORITHMS & THE DISCOVERY OF APPLICABLE ARTIFICIAL NEURAL NETWORK STRUCTURE |
| Principal Investigator: |
East, Dr IR |
| Other Investigators: |
|
| Researcher Co-Investigators: |
|
| Project Partners: |
|
| Department: |
School of Computing |
| Organisation: |
University of Buckingham |
| Scheme: |
Standard Research (Pre-FEC) |
| Starts: |
01 July 1993 |
Ends: |
30 September 1995 |
Value (£): |
87,182
|
| EPSRC Research Topic Classifications: |
|
| EPSRC Industrial Sector Classifications: |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary on Grant Application Form |
|
1.Investigate and determine effective abstract (developmental) representations of Artificial Neural Network (ANN) structure suitable for optimisation by Genetic Algorithm for practical applications. Both pure feedforward ANNs (Multiple-Layer Perceptrons) and those incorporating feedback (Dynamical Neural Networks) will be addressed.2. Investigate the properties of Diffusion Genetic Algorithms including the effects of isolation scaling and non-deterministic phenogenesis.Progress:In order to achieve acceptable performance, training rate, and generalisation, in an ANN, it is essential to discover structure appropriate to the target application. An earlier project, funded by SERC, successfully demonstrated that Genetic Algorithms offer a practicable, effective means of discovery of such structure for ANNs of significant complexity. It also provided hard evidence that Distributed Population Genetic Algorithms (DPGAs) of significant advantages in performance and scalability of parallel implementation. The purpose of this project is to seek to understand the origins of these benefits and to remove the strict limit to the complexity of ANN structure dictated by direct genetic representation. A thorough review of available theory regarding the design of genetic representation lead to the conclusion that it was wholly inadequate and misleading. Considerable effort has been expended to rectify this problem, leading to four highly significant publications and a much improved understanding of the subject. A novel approach is being taken to genetic representation by a program which avoids problems inherent in the arbitrary definition of programming language, as done by earlier workers, such as the predisposition of search trajectories. Progress towards the demonstration of the method, in an application other than ANN structure synthesis, is nearing completion. One paper has already been submitted. Progress towards the application of the method to ANN structure optimisation is significant. Completion is expected by April 1995. Implementation of Diffusion DPGAs, and developmental processes has so far been on a high performance Apple Macintosh. The availability of T9000 (much more cost-effective than the earlier T800) transputer networks has been greatly delayed. Development of a new, massively scalable, occam DPGA implementation is under way (using a Parsys SN9400SP platform) and will be complete by April 1995, when delivery of acceptable T9000s is expected. The performance thus rendered available is expected to permit convincing demonstration of the power of the new representations.
|
|
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: |
|