Descending epsilon in back-propagation: a technique for better generalization

Epsilon technique cannot learn from training patterns with a contradiction | i.e. two or more training patterns have identical inputs but dierent outputs. These training patterns are not 100% satisable and the Descending Epsilon technique will not be able to reduce the value of epsilon beyond a limit. Secondly, the Descending Epsilon technique takes more cycles than normal back-propagation since it requires more rigorous conditions for a back-propagation to occur and it has do it many times for dierent values of epsilon. Both of these problems may be eased by relaxing the condition for decreasing the value of epsilon by requiring that 90% of the errors fall within the value of epsilon and the other errors be within the previous value of epsilon. However, such a relaxation might degrade the correctness ratio and generalization. Despite these problems, our experiments show Descending Epsilon improves back-propagation with respect to satisfactory completion of training, higher correctness ratios, and improved generalization to novel examples.