Many real-world optimisation problems are subjected to dynamism and uncertainties that are often impossible to avoid in practice. The presence of dynamism and uncertainty makes optimisation problems challenging to solve by traditional optimisation methodologies like mathematic programming, especially when these problems are nondifferentiable, multimodal or highly nonlinear. A promising approach to address such optimisation problems is the use of computational intelligence (CI), which has made it possible to handle complex optimisation problems that were intractable in the past. CI encompasses a wide branch of research areas, which applies ideas from nature and biology to solve optimisation and search problems. CI has been successfully applied to many real-world scenarios where a large amount of uncertain, imprecise, and incomplete information needs to be processed. The purpose of this special issue is to showcase the latest developments in the application of CI technology to optimisation problems in dynamic and uncertain environments. The main content includes but is not limited to environment benchmarking, dynamism handling, advanced CI, theoretical analysis, real-world applications, implementation framework and software development, etc. These studies represent the trend of research on CI for dynamic and uncertain optimisation problems, with the ultimate goal of maximising CI applicability in complex application scenarios. We received an abundance of submissions from around the world, and after a rigorous peer review process, we have finally selected 8 papers for publication. The selected collection of papers covers various fascinating research topics, all of which have achieved key breakthroughs in their respective fields. We believe that these accepted papers have guiding significance for their research fields and can help researchers enhance their understanding of current trends. Sincere thanks to the authors who chose our platform and all the staff who provided assistance for the publication of these papers. In the article ‘Considering Spatiotemporal Evolutionary Information in Dynamic Multi-objective Optimisation’, Fan et al. address the challenges of preserving population diversity and providing knowledge in dynamic multi-objective optimisation (DMO) where sampling spaces vary over time and space. They propose a sliding-time-window-based population clustering method (SPC) that saves historical data for temporal knowledge and employs spectral clustering to divide data into neighbourhood subspaces for spatial diversity. Simulation results on 14 DMOPs from IEEE CEC2018, along with an actual immersed tunnel element control problem, demonstrate that SPC enhances tracking performance and reduces computational costs compared to state-of-the-art dynamic multi-objective evolutionary algorithms. In the research entitled ‘Domain-adapted driving scene understanding with uncertainty-aware and diversified generative adversarial networks’, Hua et al. tackle the domain shift problem where models trained on synthetic data from driving simulators fail to generalise to real-world autonomous driving scenarios. They propose a novel uncertainty-aware generative ensemble method that obtains complementary predictions from different optimisation objectives, training iterations, and network initialisations. By developing an uncertainty-aware ensemble scheme to fuse these predictions, the framework successfully leverages the strengths of ensembles to enhance adapted segmentation performance across three large-scale datasets. In the paper ‘Evolutionary Dynamic Multiobjective Optimisation Assisted by Inverse Regression Tree Predictor’, Gao et al. introduce a new DMOEA based on an inverse regression tree (IRT), called IRT-MOEA/D, to overcome the nonlinear relationship between objectives and decision variables. The method constructs a predictor by regression training the mapping from objective space back to decision space, allowing for the prediction of high-quality initial populations from sampled expected solutions. Statistical results on the DF series test suites indicate that this approach effectively captures the dynamic nature of complex DMOPs by utilising historical information more efficiently than existing inverse models. In the article ‘Combining Kernelised Autoencoding and Centroid Prediction for Dynamic Multi-objective Optimisation’, Hou et al. propose KAEP, a unified paradigm that reacts to environmental changes by generating two distinct subpopulations. The first subpopulation is created using a centroid-based prediction strategy, while the second employs a kernel autoencoder to predict the movement of Pareto-optimal solutions based on historical elite data. Empirical results demonstrate the superiority of this combination strategy in maintaining both convergence and diversity across a range of complex benchmark problems. In the paper ‘A Safe Reinforcement Learning Approach for Autonomous Navigation of Mobile Robots in Dynamic Environments', Zhou et al. present Conflict-Averse Safe Reinforcement Learning (CASRL) to resolve the tension between navigation efficiency and motion safety in crowded settings. The algorithm separates collision avoidance from goal-reaching tasks, maintaining a safety critic to evaluate action risks and utilising model-agnostic policy gradients to eliminate mutual interference. Extensive validation shows an average 8.2% performance improvement over vanilla baselines in simulated dynamic environments, with successful deployment confirmed through forty real-world robot experiments. In the research entitled ‘AI-enabled Bumpless Transfer Control Strategy for Legged Robot with Hybrid Energy Storage System’, Huang et al. propose an intelligent control strategy to eliminate torque bumps during mode switching in legged robots. The framework integrates a Proximal Policy Optimisation (PPO)-based non-linear active disturbance rejection controller with a Deep Neural Network (DNN)-based bumpless transfer strategy to set initial controller values at switching moments. Simulations and experimental results validate that this AI-enabled approach effectively enhances stability and energy efficiency by smoothing the transition between driving and regenerative braking modes. In the paper ‘MOEA/D-based multi-row facility layout optimisation method with discontinuity perceiving of the Pareto front’, a novel multi-row facility layout problem (MRFLP) is formulated to minimise both layout area and flow cost, specifically accounting for the orientation of input/output points. Guo et al. introduce a decomposition-based evolutionary algorithm that perceives the discontinuity of the Pareto front and adjusts weight vectors to enhance exploitation in continuous regions. By incorporating a re-initialisation mechanism and an improved adaptive epsilon constraint-handling technique, the method effectively balances diversity and feasibility, outperforming other state-of-the-art algorithms across 12 test instances. In the article ‘Building Blocks as Experiences in Dynamic Capacitated Arc Routing Problems’, Tong et al. propose a dynamic optimisation framework with a solution building block adaptation strategy (DO-SBBA) to reuse experience from previous task instances. Unlike traditional methods that re-optimise from scratch, this framework extracts “building blocks” from historical solutions to facilitate the optimisation of new instances with cost- or task-changing events. Experimental studies confirm that DO-SBBA significantly improves performance in dynamic scenarios by effectively leveraging valuable historical optimisation experience. We express our gratitude and congratulations to all authors of papers published on this special issue of CAAI transactions on intelligence technology. Their high-quality research will strongly promote the development of related fields. We would also like to thank all the reviewers who participated in the peer review of these papers, as well as the chief editor and editorial office of CAAI Transactions on Intelligence Technology, for their support throughout the entire process. We gratefully acknowledge the contributions of our fellow guest editors: Shengxiang Yang, Juan Zou, Changhe Li to this special section. Research data are not shared.
Shouyong Jiang (Mon,) studied this question.