Xiaoming Xue (薛小明)

Xiaoming Xue is currently a Postdoctoral Fellow with the Department of Data Science and Artificial Intelligence (DSAI) at The Hong Kong Polytechnic University, working with Prof. Kay Chen Tan and Dr. Jibin Wu, as a member of the Machine Intelligence and Nature-Inspired Computing (MIND) Lab. He received his B.Eng. and M.Eng. degrees from China University of Petroleum (East China) under the supervision of Prof. Kai Zhang, and the Ph.D. degree from City University of Hong Kong under the supervision of Prof. Kay Chen Tan and Prof. Linqi Song.

His research interests include transfer optimization, surrogate modeling, evolutionary computation, machine learning, and their practical applications in real-world scenarios.

Email  /  Google Scholar  /  Github  /  CV

Selected Publications (Including Preprints)

Theories

This paper aims to establish theoretical foundations for analogy-based evolutionary transfer optimization (ETO), with a particular focus on supporting various algorithms that rely heavily on a key concept known as similarity. We first connect analogical reasoning to ETO's key issues—what, how, and when to transfer. Next, we develop theories for analogy-based knowledge transfer, including three ideal knowledge transfer functions and two theorems addressing positive and negative transfers. Finally, we provide a novel insight into ETO's conditional superiority over traditional evolutionary optimization through the lens of the no free lunch theorem.

Algorithms

Both the optimized solutions from source tasks and the promising solutions acquired by the target task are treated as task-solving knowledge, enabling them to compete with each other to elect the winner for expensive evaluation, thus boosting the search speed on the target task. Moreover, the nonnegative performance gain resulting from such competitive knowledge transfer is analyzed theoretically.

By dividing the optimization experience into two categories, i.e., global and local search experience, a novel transferability metric named landscape encoding-based rank correlation is proposed and instantiated for global and local transfer, which can work cooperatively for maximum performance enhancement for the target task.

Given a knowledge base with single-objective optimization problems (SOPs), multi-objective optimization problems (MOPs), and many-objective optimization problems (MaOPs), we propose to divide past search experience into two categories based on decision variable analysis: convergence-related and diversity-related experience. Subsequently, two distinct knowledge transfer methods are developed to utilize these two types of search experience for better addressing the convergence-related and diversity-related variables of a target task of interest, respectively.

To improve the transferability of solutions for heterogeneous multi-task optimization problems with effective inter-task mapping, a novel rank loss function is proposed to acquire superior inter-task mappings. Specifically, using an evolutionary-path-based representation model for optimization instances, an analytical solution for affine transformation is mathematically derived to bridge the gap between two distinct problems, which can be seamlessly embedded into any evolutionary multi-task solver.

An efficient single-pool decomposition framework has been developed to identify the interactions among large-scale decision variables in an ordinal fashion. Additionally, it has been found that the decomposition efficiency can be greatly improved by leveraging the variable-related topological information of the problems of interest. In many real-world scenarios, this information could be 1-, 2-, or 3-dimensional coordinates, which represent the geometric structure of large-scale systems.

Benchmarks

A scalable generator for sequential transfer optimization problems (STOPs) with customizable source-target similarities with respect to optimum is developed to enable more comprehensive evaluation of various analogy-based algorithms transferring optimized solutions. Rationale: In analogy, decisions about knowledge transfer are made based on similarities.

In the case of simply connected Pareto manifolds (PMs), the universal representation of source-target PM similarities is achieved through a bottom-up approach based on homeomorphisms, allowing for a close resemblance to the diverse PM similarities found in real-world multiobjective sequential transfer optimization problems.

Surveys

A review of the recent advances in evolutionary multiform transfer optimization is conducted from the perspectives of the problems being solved and the methods used to construct auxiliary formulations.

A holistic review of various techniques associated with the fundamental aspects of solution transfer (i.e., what, when and how to transfer) is presented. Additionally, a series of experiments are conducted to explore the underlying transfer mechanisms that enhance the evolutionary search with diverse source-target similarities in the context of continuous optimization with sequential transfer.

Applications

A competition is initiated between the production schemes from the source reservoirs and the target reservoir. The winning scheme is then selected for reservoir simulation to evaluate its quality, enabling the identification of a superior production scheme while reducing the need for extensive reservoir simulations. Additionally, the non-negative gain of the proposed algorithm is highlighted, promoting a more reliable application of this technology in oilfields.

A clustering-based similarity measurement method for evolutionary sequential transfer optimization is developed to accurately select promising source solutions to speed up the evolutionary search in solving trajectory optimization problems.

This work is the first attempt to propose a semi-supervised transfer learning framework dubbed S3TL for CAD of mental disorders using fMRI data, in which a cross-domain feature alignment method is developed first to generate target-related source model. Then, an enhanced dual-stage pseudo-labeling method is developed to assign pseudo-labels for unlabeled target samples. Lastly, a knowledge transfer method is used to improve the generalization capability of the target model.

By modeling production optimization as a dynamic optimization problem consisting of a series of time windows, the optimization data from historical windows are used as part of the training data to construct the transfer Gaussian process (TGP) for effectively solving the current expensive optimization task. This approach avoids the extensive time required for "trial and error" and achieves superior performance, similar to that of experienced engineers.

Given a large-scale production optimization problem, one may choose to decompose it into a number of simpler, lower-dimensional subproblems. Then, multiple data-driven surrogates are built for these subproblems. Finally, all the subproblem surrogates are optimized cooperatively using a reuse strategy for subproblem samples.

A novel self-adaptive multifactorial evolutionary algorithm (SA-MFEA) is developed for multi-task production optimization problems, with the transfer intensity estimated online based on the representation models of optimization tasks to reduce the risk of negative transfer. In SA-MFEA, the positive transfer between highly related tasks can be greatly enhanced by estimating a higher transfer intensity, while the negative transfer between dissimilar tasks can be minimized with a low value of transfer intensity.

A multi-fidelity genetic transfer framework is developed for production optimization problems, wherein a number of low-fidelity formulations are available to speed up the search for the high-fidelity task. Particularly, a multi-fidelity transfer differential evolution (MTDE) algorithm is developed by incorporating a novel two-mode genetic transfer strategy into differential evolution. The efficacy of MTDE was validated using the egg model and two real field cases, where the black-oil and streamline models are employed to obtain high- and low-fidelity results, respectively.

A global and local surrogate-assisted differential evolution (GLSADE) algorithm is introduced for waterflooding production optimization problems. Specifically, GLSADE consists of two phases: 1) a global optimization phase that generates multiple offspring, and 2) a local search phase that seeks the optimum of the surrogate. The cooperation between these two phases significantly enhances the optimization process in terms of search speed and final solution quality.

Dissertation

This thesis focuses on consolidating the theoretical foundations of sequential transfer optimization in the context of evolutionary computation and developing relevant algorithms based on these foundations, which is structured around a research path that begins with underlying principle, progresses through theoretical foundations for knowledge transfer, and concludes with problem and algorithm designs.

Honors and Awards

Academic Service

Journal Reviewers

• IEEE Transactions on Evolutionary Computation (TEVC)
• IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
• IEEE Transactions on Neural Networks and Learning Systems (TNNLS)
• IEEE Transactions on Cybernetics (TCYB)
• IEEE Computational Intelligence Magazine (CIM)
• IEEE Transactions on Cognitive and Developmental Systems (TCDS)
• IEEE Transactions on Emerging Topics in Computational Intelligence (TETCI)
• IEEE Transactions on Artificial Intelligence (TAI)
• Nature Communications (NC)
• Artificial Intelligence Review (AIR)
• npj Artificial Intelligence (npj AI)
• Engineering Applications of Artificial Intelligence (EAAI)
• Applied Soft Computing (ASOC)
• Cognitive Computation (COGN)

Conference Reviewers

• IEEE Congress on Evolutionary Computation (CEC)
• International Joint Conference on Neural Networks (IJCNN)
• IEEE Conference on Artificial Intelligence (CAI)
• Annual AAAI Conference on Artificial Intelligence (AAAI)

This website's source code is adapted from here.