Du, Ke-Jing (2024) Advanced Evolutionary Computation For Dynamic And Multi-Task Optimization Via Efficient Knowledge Transfer. PhD thesis, Victoria University.

Abstract

Evolutionary computation (EC) is a kind of population-based search method, drawing inspiration from natural selection and gene inheritance. Although EC has shown advantages over traditionally mathmetic-based optimization methods, it often neglects a crucial aspect: knowledge gained from past and other problem-solving experiences. This thesis explores how EC can improve by learning from past experiences or experiences across different tasks. Inspired by knowledge transfer (KT) observed in human evolution through cultural genes, this thesis investigates the potential of EC to acquire and apply knowledge from past and other problem-solving experiences and focuses on dynamic optimization problems (DOP) and multi-task optimization problems (MTOP). Because these types of problems offer ideal opportunities for KT. DOP involves dynamic changes over time, while MTOP optimizes multiple tasks simultaneously, both scenarios benefiting from problem-solving experiences. This thesis emphasizes the importance of KT, proposing novel EC algorithms for efficiently solving DOP and MTOP. In DOP, a challenge lies in effectively utilizing historical information to accelerate algorithm convergence. This requires solutions for selecting and updating historical data and ensuring its validity amidst environmental changes. Meanwhile, MTOP presents difficulties in balancing optimization objectives across multiple tasks and designing appropriate differential evolution strategies to accommodate different task properties and constraints. Real-world problems are often more complex than theoretical research, involving numerous variables and factors that leading to dynamic charactertics and multi-task charactertics. For example, in bike-sharing systems, fluctuations in the numbers of available bicycles and stations make the path planning a DOP. Another example is feature selection in deep learning for high-dimensional data. Due to the curse of dimensionality, considering all features is challenging. Thus, integration of KT and EC is essential for addressing practical problems. Main contents and contributions of this thesis are detailed as follows. 1. To address DOP efficiently, a historical information-based differential evolution (HIDE) algorithm is proposed in Chapter 3. HIDE uses previous knowledge for faster convergence to new optimal regions and employs an archive-based strategy to retain the best-performing individuals from previous environment, facilitating effective KT. 2. To tackle MTOP, a multi-criteria EC algorithm is proposed in Chapter 4. MTOP is conceptualized as a multi-criteria optimization problem (MCOP), where KT occurs across a consolidated population. 3. Chapter 5 addresses dynamic user route planning problem (URPP) in bike-sharing systems. The challenge of fluctuating station inventory turns URPP into DOP. To utilize experiential knowledge and guide the algorithm’s search process, knowledge learning and random pruning-based memetic algorithm (KLRP-MA) is introduced, enhancing KT integration and effectively tackling URPP dynamics. 4. In Chapter 6, a practical application of integrating MTOP with the bi-directional feature fixation (BDFF) method in multi-tasking bi-directional particle swarm optimization (MBDPSO) is discussed. This integration allows for effective KT between tasks, improving capabilities in high-dimensional feature selection for deep learning. In summary, this thesis systematically investigates DOP and MTOP and their practical applications, proposing advanced EC algorithms to enhance the efficiency of KT.

Search Google Scholar

Repository staff login