Relearning and Evolution in Neural Networks

this paper (Parisi, Nolfi, & Cecconi, 1992). The performance of the elite did not improve when lifetime learning of the second task was introduced, whereas average performance did improve. It seems clear that the effect of lifetime learning was merely to go some way towards restoring performance of networks which had had their weights perturbed (by mutation) away from trained (through evolution) values --- a form of relearning. The extreme convergence of the population around the clustered elite members of the previous generation should be borne in mind when reading from (Nolfi et al., 1994), p. 22: The offspring of a reproducing individual occupy initial positions in weight space that are deviations (due to mutations) from the position occupied by their parent at birth (i.e., prior to learning). One form of relearning in networks was analysed in (Hinton & Plaut, 1987). In that case a network is first trained by some learning algorithm on a set of input/output pairs; the weights are then perturbed. After retraining on a subset of the original training set, it is found that performance improves also on the balance of the original training set. The present case differs from this, in that the lifetime learning is on a fresh task, rather than on a subset of the original task. Recently just such an effect was predicted and observed in networks (Harvey & Stone, 1995). When good performance on one task is degraded by random perturbations of the weights, then in general training on any unrelated second task can be expected to improve, at least initially, the performance on the first task. C P Q B B A 1 2 Figure 1: A two-dimensional sketch of weight space. To briefly summarise the reasons for this, consider the diagram, which represents the weight space of a network in just 2 ..