A Generic Recommender System for Continuous Optimization based on Deep Neural Network

Abstract

As per the no lunch theorem it is revealed that No one algorithm can exceed anybody else on all classes of optimization. To tackle this problem, methods have been developed to recommend an existing problem solving algorithm. However, there is poor practicality and transferability of existing recommendation methods for contentious optimization, mainly because of the problem of extracting features which can effectively describe the structure of the problem and lack of training data. This paper proposes a generic system to deal with the two challenges mentioned above.

First, it is proposed a new method to represent the analytical objective function of a tree that is used directly as a continuous optimization problem. Second, on the basis of the proposed tree representation, a large number of benchmark problems are created randomly providing a large number of training data with different levels of difficulty. A recommendation model is trained by a deep recurrence network in which a metaheuristic algorithm for optimization of white or black box is recommended, which a significant step is towards fully automated recommendation for continuous optimization of algorithms. This paper proposes a system that can automatically select a metaheuristic algorithm that is suited to a particular problem without trial or error. The method suggested Develops a tree-like data structure to display the difficulty in optimizing and training a deep recurrence neural network to learn the best metaheuristic solution.