An Evolutionary Approach to Designing Neural Networks
Abstract
One of the most interesting properties of neural networks is their ability to learn appropriate behavior by being trained on examples. Established learning algorithms, which typically work by minimizing error through backpropagation in weight space, tend to get stuck in local optima--a tendency typical of gradient-descent methods applied to nonconvex objective functions. Therefore, for problems of nontrivial complexity these systems must be handcrafted to a significant degree, but the distributed nature of neural network representations make this handcrafting difficult. We are investigating an evolutionary approach to learning that will avoid this problem. This approach simulates a variable population of networks which, through processes of mutation, combination, selection, and differential reproduction, converges to a group of networks well suited to solving the task at hand. The important components of the approach are a genetic language for coding a large variety of networks, a procedure for constructing networks from these genetic codes, a nondeterministic process for mutating and combining genetic codes, and a function that measures the overall fitness of networks in the context fo the task at hand. We use a Connection Machine to exploit the inherent parallelism in the simulations.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1990
- Accession Number
- ADA228874
Entities
People
- Aviv Bergman
- Stephen Barnard
Organizations
- SRI International