A new tool for market research using a modified auto-associative memory

This paper describes a new application of the Hopfield auto-associative memory. This kind of network has been used as a very good tool for solving optimization problems, but not as a tool for the investigation of public opinion polls, market research or stock market predictions. We have considered the training set as a sample of public-opinion polls. The aim of this research is to design the net as a tool for obtaining statistical information on the general opinion of the population represented by that sample. To do this, based on preliminary research, we used a modified version of the Hopfield auto-associative memory. Once the weights that control the net have been obtained, we may, through a set of “external parameters”, control the number of fixed points of the net. Moreover, the state vector space may be classified in as many classes as the dimension of the space and we can control the number of fixed points in any one of those classes. So, not only do we have the degree of correlation between a given element of the population and the sample of polls, but we also have the degree of correlation between that element and any one of the classes of the space. Associated with any element of the population, we also have a probability frequency distribution which may be used to obtain the ratio of elements of the population whose answer to a certain question is similar to the one given by the element to be tested. We present the general features of the theory and prove the performance of the algorithm