Exactness and Component Sharing in Expensive Evolutionary Multiobjective Optimization

Description

Many real-world applications involve optimizing multiple costly-to-evaluate conflicting objectives, such as finding strong yet ductile material, building a neural network with high accuracy and low latency, and designing products with high quality and low cost. Many algorithms have been proposed to produce a finite set of different trade-off solutions for decision makers to choose their preferred ones. Existing algorithms have two major drawbacks: (i) A finite set of solutions may not contain the ones that exactly match the decision maker's preferences. A large amount of computational resources is indeed wasted on searching for those not-very-useful solutions. (ii) In many applications, a preference could correspond to a different version of product or application scenario. To reduce the manufacturing cost and support module designs, it usually requires that different optimal solutions should share some common components. No existing expensive multiobjective optimization algorithm consider this requirement. This project makes a first systematic attempt to address the exactness and component sharing issues in expensive multiobjective optimization. We will develop and study Bayesian optimization algorithms for finding a finite number of optimal solutions to exactly fit the decision maker's preferences, a linear function for modelling all the optimal solutions close to a given preference, and a general model for all the optimal solutions to an expensive multiobjective optimization problem. All these algorithms treat component sharing as their constraints. This project will make the best use of ideas and techniques from evolutionary computation, single objective expensive optimization algorithms, traditional optimization, and machine learning. The proposed algorithms will be tested on shape optimization of wind turbine blades.