A Study Of High Performance Multi-task Intelligent Opimization Algorithm Via Knowledge Transfer

Intelligent optimization algorithms,such as simulated annealing algorithms,particle swarm optimization algorithms,and evolutionary algorithms,are a series of heuristic algorithms that have been inspired by nature to solve complex optimization problems.In contrast to the classical numerical optimization methods,such as gradient descent and Newton method,intelligent optimization algorithms have week dependence on the properties of objective functions,strong global-search capabity to avoid falling into local optimum and wideapread application background in real-world.However,as traditional intelligent optimization algorithms solve only one problem in a single run,they can not make use of the similarity among the problems,which could lead to improved optimization performance when properly harnessed.Different from traditional intelligent optimization algorithms,multi-task intelligent optimization aims to solve multiple different optimization problems simultaneously.It intendes to achieve enhanced optimization performance through knowledge transfer among tasks.The emergence of multi-task intelligent optimization is partially motivated by the rapid development of cloud computing.Solving multiple different tasks at the same time is the basic functional requirement of the cloud computing platform.Multi-task optimization algorithms fit well in the cloud computing framework to provide technical support for efficient optimization.However,as a new research topic in the field of intelligent optimization,research progress in the field of multi-task optimization is still in its infancy.There are still many problems to be explored and solved.In this thesis,we focus on the inter-task knowledge transfer in multi-task optimization to develop high-performance multi-task optimization algorithms.The main contributions are concluded as follows:(1)We investigate the impact of different crossovers for conducting knowledge transfer in multitask optimization algorithm.Further,towards robust multitasking performance on different optimization tasks,we propose a novel multitask optimization algorithm with adaptive knowledge transfer.In the proposed algorithm,a number of crossover operators with unique search bias are adopted and the crossover operator for knowledge transfer across tasks is configured automatically based on the information collected while the evolutionary search progresses online.Owing to the adaptive design,the proposed algorithm achieved good performance on different multi-task optimization problems.(2)We introduce a denoising autoencoder-based explicit knowledge transfer method and further propose a new multi-task optimization framework.In this framework,each task has an independent solution representation and evolutionary solvers with different biases are employed for solving different tasks simultaneously.The algorithm uses a denoising autoencoder for conducting knowledge transfer,in the form of problem solutions,across tasks explicitly.The explicit knowledge transfer method,as well as the employment of multiple evolution solvers for multitasking,greatly improved the performance of the multi-task optimization algorithm.(3)In view of the participation of private cars in customer service,we first present a vehicle routing problem with heterogeneous capacity,time window and occasional driver(VRPHCTWOD)and construct the VRPHCTWOD benchmark.Futher,we propose a permutation-based multi-task optimization algorithm to solve multiple VRPHCTWODs simultaneously.Lastly,we use the VRPHCTWOD as a case study to analyze and investigate the performance of multi-task optimization on combinatorial optimization problems.