Abstract
This paper proposes Intersection Learning algorithm for Bidirectional Associative Memory (ILBAM). The condition which guarantees the recall of all training data in the BAMs is given by a system of linear inequalities. In order to solve the system of linear inequalities, some learning algorithms using relaxation methods have been proposed. However, the weight renewal times of the algorithms depend on the correlation of training data. In this paper, we derive a relaxation method based on geometrical consideration and apply it to the learning of the BAMs. A number of computer simulations show the following effectiveness of the proposed ILBAM algorithm: (1) It can guarantee the recall of all training data. (2) It requires much less weight renewal times than the conventional methods. (3) It becomes more effective in case there are many training data to be stored. (4) It is insensitive to the correlation of training data. (5) It contributes to the noise reduction effect of the BAMs.
