Improved Alopex-based Evolutionary Algorithms And Their Applications To Process Modeling

Alopex-based Evolutionary Algorithm (AEA) has the dual characteristics of simulated annealing (SA) and gradient descent, utilizing heuristic correlation information of the swarm. However, its application to complicated function has exposed its shortcomings of trapping in local optimum.In this paper, the AEA is improved by three different strategies to overcome its shortages, named as NAEA, GAEA and LAEA. NAEA does well in the functions with weakly-correlated variables. GAEA improves the NAEA by a Gaussian Copula function, which is suitable for the functions with strong-correlated variables, with features of both fast convergence and high convergence precision. LAEA introduces a local search mechanism to keep the diversity of the population, so as to improve the success rate. Numerical simulation results of18benchmark functions and applications to chemical processes have indicated their effectiveness and usability. Then, a new strategy model that is proposed to avoid the over-fitting problem in Back-Propagation Neural Network (BPNN) model for the situation of modeling a system with a small quantity of samples. This approach is based on BPNN model combined with monotonicity and concavity knowledge and LAEA. In this new strategy model the monotonicity and concavity knowledge gain from expert knowledge about the process is used as constraint conditions to the objective function of neural network. Thus, the training process is changed into a constraint problem optimization. Finally, a new evolutionary algorithm named LAEA is used to search the optimal weights and thresholds. Furthermore, we used this new strategy model to the over-fitting simulation of sine function and the cracking furnace propylene yield, and the satisfactory result shows its suitability for practical applications.