ON THE DISTRIBUTION OF PERFORMANCE FROM MULTIPLE NEURAL-NETWORK TRIALS

The performance of neural-network simulations is often reported in ter ms of the mean and standard deviation of a number of simulations perfo rmed with different starting conditions, However, in many cases, the d istribution of the individual results does not approximate a Gaussian distribution, may not be symmetric, and may be multimodal, We present the distribution of results for practical problems and show that assum ing Gaussian distributions can significantly affect the interpretation of results, especially those of comparison studies, For a controlled task which we consider, we find that the distribution of performance i s skewed toward better performance for smoother target functions and s kewed toward worse performance for more complex target functions, We p ropose new guidelines for reporting performance which provide more inf ormation about the actual distribution.