特異領域を活用した多層パーセプトロン探索法

摘要

In a search space of multilayer perceptron having J hidden units, MLP(J), there exists a singular flat region created by the reducibility mapping of the optimal solution of MLP(J-1). Since such a singular region causes serious slowdown for learning methods, methods for avoiding or eliminating the region have been proposed, but have failed to achieve their goals so far. This paper proposes a new learning method which makes good use of singular regions to find successively MLP solutions good enough for each J. The method is evaluated by experiments using artificial and real data.%多層パーセプトロンの探索空間には，隠れユニット数が一つ少ないモデルの最適解と等価な出力を出すパラメータ領域が特異領域を形成し，学習の停滞を招くため回避や解消が模索された.本稿では，可約性写像を用いて特異領域を積極的に利用して，探索空間に多数存在すると思われる局所最適解の中から，さまざまな隠れユニット数にふさわしい良解を系統的に得る探索法(特異階段追跡法)を提案し，計算機実験にて有効性を評価する.