This paper gives a new learning rule about the formation of weights for two-layer max-min feedforward fuzzy associative memory (FAM) network proposed by Kosko . Based on the new rule

The feedforward FAM model is developed into a fuzzy bidirectional associative memory (BAM) model

and a fuzzy quick augmentation algorithm is also proposed

Its stability and tolerance for the BAM model are also analyzed. From the analysis

an interesting result which can store an arbitrary given multi-value patterns is obtained. When used to store binary values

The weights for BAM model take binary too

0 or 1.So it is suitable for the VLSI and optical implementation. In order to make a comparision

binary based sample patterns have adoped. A larger number of simulation results show the advantages of a less number of weighted value

or the simple implementation

by comparing with the existing learning algorithm

such as binary based Hoperfield dummy augmentation and MBDS augmentation algorithms. On the other hand

the fuzzy quick augmentation algirithm has the merit of the simpler computation and faster convergence.