Directional cooperative networks

Abstract

Although various machine learning methods have been proposed for industrial applications, there are not many examples of their application in some industrial sectors. Data collected within organizations are inconsistently formatted and expensive to annotate, resulting in insufficient datasets for training. This hinders the widespread use of machine learning methods. The challenge is to address the need for more performance of machine learning models due to the lack of available data.We address this problem by introducing Directed Cooperative Networks (DCNs). This approach addresses the lack of data by connecting two neural networks with a function that evaluates the output of Generator. Estimator, the one of the networks, acts as an approximate function of the evaluation function, assisting Generator to output a product that yields the desired evaluation function value. When the networks are sufficiently trained, Estimator’s output approaches the output of the evaluation function, and Generator will produce products with the desired attributes.A data-independent method is the evolutionary algorithm (EA). Since EA is optimized by feedback from the environment, good results can be obtained by using the evaluation function of the product as its environment. However, its effectiveness decreases as the combination of product components becomes more complex.One of the outstanding properties of DCN is its potential to outperform EA on complex problems because it learns using the gradient of the search field. DCN does not require rewriting the evaluation function into a back-propagatable form, even if the function is complex. Using a neural network that behaves as an approximate function of the evaluation function, the gradient descent method can be applied even if the evaluation function is non-differentiable.To validate the effectiveness of the proposed method, DCN is applied to a molecular search task and analyzed in comparison with other approaches. This study aims to demonstrate the superior performance of DCN in dealing with data shortages in complex problems in industrial applications.