Research

Participated in research of the following areas: evolutionary computation, computational optimization, machine learning, data analysis, numerical analysis, parallel algorithms.

Research Interest

• • Theoretical analysis of evolutionary algorithms and other randomised search heuristics

  • Multi-objective evolutionary algorithms for constrained optimization

  • Coevolutionary system, game strategy and artificial life

  • Applications of machine learning in cyber-security, e.g. networks intrusion detection

  • Applications of machine learning in data analysis, e.g. electronic health records

  • Applications of artificial intelligence in language education

Research Highlight

• 1. Drift analysis of evolutionary algorithms and randomised search heuristics (He & Yao 2001, He & Yao 2017). 

     • "Drift analysis was introduced to the theory of evolutionary algorithms by He and Yao. It soon became one of the strongest tools both for proving run-time guarantees for many evolutionary algorithms and for giving evidence that some algorithms cannot solve certain problems.'' 

     • "Drift analysis, which is widely used by now, has been put forward by He and Yao.'' 

     • "drift analysis, a method that provided important insights into the computational complexity of discrete EAs over the last decade.'' 

     • "Expected runtime analysis inspects the average runtime of an algorithm on a particular problem, and can exploit mature probabilistic techniques, such as drift analysis and others.'' 

     • "drift analysis, a standard tool in theory of randomised search heuristics."

     • "Drift analysis is one of the major tools for analysing evolutionary algorithms and nature-inspired search heuristics."

     • "Moreover, the well-known drift analysis has been introduced by He and Yao (2001, 2003), serving as a general theoretical framework. It was employed for several challenging theoretical questions."

     • "Drift analysis is a very general mathematical technique that allows to make statements regarding the long term performance (i.e., the runtime) of a randomised search heuristic by analysing the expected progress that the algorithm makes in one step."

     • "The additive drift theorem is commonly used to derive upper (resp. lower) bounds on the expected runtime of an algorithm from lower (resp. upper) bounds on the expected increase of a suitable potential. It first appeared in the analysis of evolutionary algorithms in He and Yao's work [60], [61], in which they implicitly used the optional stopping theorem for martingales. "

     • "In this paper, we use the drift analysis, in particular the average drift analysis [15], to estimate the EHT of the ENAS algorithms."

  2. Helper and equivalent objective differential evolution for constrained optimization (HECO-DE) (Xu, He & Shang 2022)

     • HECO-DE was ranked 1st in 2019 in IEEE CEC Competition on Constrained Real Parameter Optimization.

  3. Average convergence rate of evolutionary algorithms (He & Lin 2016,  Chen & He 2021).

     • "The fourth type of convergence representation is the average convergence rate that specifically measures the rate of fitness change as proposed by He and Lin."

     • "Convergence ratio can be evaluated by Markov chain analysis, however, this process is complicated from theoretical and practical point of view.  For this reason, the convergence indicator is described as (He and Lin, 2016)." 

     • hence the asymptotic convergence rate constant is defined through the geometric average of the one step decrease ratios, [8]. 

     • The average convergence rate (He and Lin, 2015, Chen and He, 2019, Tuo et al., 2020) is used as a measure of convergence. 

     • "In this work, Average Convergence Rate (ACR) method is utilized to analyze the convergence rate of BOA and EBOA."

     • "The alternative average convergence rate (AACR) analysis is also performed to visualize the convergence speed and stability of the proposed DARC-DE algorithm."

     • "To situate the contribution of this article within the literature, it is important to mention that in [26], Chen and He present a theoretical analysis of ACR in continuous optimization for evolutionary methods and prove that ACR is linear when the mutation is positive-adaptive but sub-linear when the mutation is invariant or zero-adaptive."

     • "Alternative average convergence rate (AACR) (Chen & He, 2021) provides a practical method for quantifying and illustrating the convergence speed of DLMDGWO"

     • "we use the average convergence rate proposed in [55] to assess the convergence speed of the DGA compared to the standard GA, noting that the addition of the pivot individual, which conducts a targeted search, indeed results in the average error reaching zero more quickly."

  4. Easiest and hardest functions to evolutionary algorithms (He & Yao 2015).  

     • "the theory in Ref. [12] indicates easy problems for one algorithm may be difficult for another, and vice versa"

     • "Recently, He et al. [14] also presented the easiest and hardest fitness functions for RSHs, and revealed that the unimodal functions are the easiest and deceptive functions are the hardest in terms of the time-fitness landscape. This further explains that RSHs are not always efficient for all problems."

Research Grant

• • 2011-2015 Principal Investigator: Evolutionary Approximation Algorithms for Optimization: Algorithm Design and Complexity Analysis. EPSRC (£331K).

  • 2005-2008 Researcher Co-investigator: Computational Complexity Analysis of Evolutionary Algorithms. EPSRC (£291K).