Optimisation of artificial neural network structure using Direct Encoding Graph Syntax (DEGS)
- Brunel Univ., Uxbridge (United Kingdom)
An artificial neural network (ANN) is intended to represent usually a complex non-linear mapping between the two data sets that can then be able to generalize on unseen data for the solution of a particular task. The evaluation of the correct ANN structure (and hence the mapping) is very often, solely a ANN and error procedure which may not lead to the required solution. The Genetic algorithm (GA) has been perceived by researchers as a effective systematic technique for the design of ANNs. However the GA can be hampered by the difficulty of generating a variety of ANN structures. In addition there is the problem of a significant increase of the search space for network architectures as the network size increases (scalability problem). Even if these problems are addressed, the ANN structures produced by the GA must be viable and then efficiently trainable by a competent training algorithm. A network is not viable if it is incomplete with isolated processing units. Also the possibility of encountering the permutation problem which refers to the creation of ANNs that are different in structure but are equivalent geometrically also has to be reduced as this significantly reduces the efficiency of the GA. The above characteristics are indicative of other encoding schemes that poorly encode the ANN. This paper describes a direct encoding scheme, Direct Encoding Graph Syntax (DEGS), that endeavors to overcome these flaws. Its successful implementation in conjunction with the GA, for the design of ANNs to evaluate the 9-bit parity problem is also discussed.
- OSTI ID:
- 501626
- Report Number(s):
- CONF-961239-; TRN: 97:002723-0045
- Resource Relation:
- Conference: IDPT-2: 2. world conference on integrated design and process technology, Austin, TX (United States), 1-4 Dec 1996; Other Information: PBD: 1996; Related Information: Is Part Of Integrated design and process technology. Volume 1; Cooke, D.; Kraemer, B.J.; Sheu, P.C.Y.; Tsai, J.P.; Mittermeir, R. [eds.]; PB: 488 p.
- Country of Publication:
- United States
- Language:
- English
Similar Records
Correlation-Informed Permutation of Qubits for Reducing Ansatz Depth in the Variational Quantum Eigensolver
An enhanced artificial neural network with a shuffled complex evolutionary global optimization with principal component analysis