Using evolution to improve neural network learning: pitfalls and solutions

Abstract

Autonomous neural network systems typically require fast learning and good generalization performance, and there is potentially a trade-off between the two.The use of evolutionary techniques to improve the learning abilities of neural network systems is now widespread. However, there are a range of different evolutionary approaches that could be applied, and no systematic investigation has been carried out to find which work best. In this paper, such an investigation is presented, and it is shown that a range of evolutionary techniques can generate high performance networks, but they often lead to unwanted side effects, such as occasional instances of very poor performance. The nature of these problems are explored further, and it is shown how the evolution of age dependent plasticities and/or the use of ensemble techniques can alleviate them. A range of techniques are thus identified, with differing properties, that can be matched to the specific requirements of each application.

Artificial Bee Colony Training of Neural Networks

“…One solution is to use an evolutionary algorithm to optimize the key BP learning parameters, such as the random initial weight range [-ρ, ρ] and learning rate η. With a fixed, sufficiently large, number of training epochs for each problem, the evolved learning rate is then able to implement a form of early stopping and avoid over-fitting, and that consistently leads to improved performances [3]. However, such evolutionary approaches tend to be rather computationally intensive, and might be regarded as giving BP an unfair advantage over the ABC.…”

“…However, such evolutionary approaches tend to be rather computationally intensive, and might be regarded as giving BP an unfair advantage over the ABC. A standard non-evolutionary approach will therefore be studied first, but using information that consistently emerges from evolutionary investigations [3], namely that very small initial weight ranges and very slow learning rates tend to work best, with a standard stopping early approach to set the number of epochs. The details of the experimental set-up and analysis were chosen to provide the closest possible match with the ABC approach discussed in the next section.…”

“…It is known that optimizing the initial random weight range for BP can have a big effect on the generalization performance [3], so it is not surprising that it also has a big effect for the ABC too. This can easily be tested by starting with the default ABC colony size of 30 and abandoning limit of 1000 used by Karaboga and Ozturk [15], and applying ABC training with a wide 7 range of different search space limits to find the best for each dataset.…”

“…When automatic optimization of such a framework has been carried out using evolutionary computation techniques, large differences in BP learning rates have emerged across the network components, leading to significant improvements in performance for some datasets [3]. It is therefore a reasonable hypothesis that this is the factor that allows the ABC to perform better than BP on the Gene dataset.…”

“…As noted above, simulated evolution provides a reliable approach for optimizing BP learning parameters [3]. This section will use that technique to search for further improvements beyond the BP parameter values used earlier.…”