A weighting approach for autoassociative memories to maximize the number of correctly stored patterns

An autoassociative memory can store multiple information coded as binary patterns in a recurrent artificial neural network (ANN), and if a similar information is presented, the most likely pattern can be retrieved. However, because some patterns can´t be stored in it correctly, we had developed a weighting approach to given patterns in order to store all of them perfectly. Nevertheless, it may be impossible to store all of them correctly so that the above method can´t always find such weights. In this paper, we propose another weighting method in order to maximize the number of stored patterns in the network. We can determine the weights by solving a mixed-integer linear programming (MILP) problem. Numerical examples demonstrate the effectiveness of the proposed method.