Sequential Transfer via Clustering-Based Similarity Measurement for Faster Trajectory Optimization

In evolutionary sequential transfer optimization (ESTO), the precision of similarity measurements between source and target tasks is crucial in selecting useful source solutions to accelerate the optimization of the target task. Despite the success of existing methods in optimizing various numerical functions, their capability to accurately measure the similarity among practical trajectory optimization problems (TOPs) remains limited due to the complex nature of real-world scenarios. To alleviate their limitations, this paper proposes a clustering-based similarity measurement method for ESTO, aiming to accurately select promising source solutions to speed up the evolutionary search in solving TOPs. Specifically, all candidate source TOPs that have been optimized are first categorized into different clusters based on their populations via k-means clustering. Afterward, the populations of source TOPs in the same cluster are employed to compute the similarity of these source TOPs to the target TOP that is being optimized. In this way, the TOPs with the highest similarity can be identified and their optimized solutions will be injected into the population of the target task for faster convergence. The experimental results demonstrate that our proposed method outperforms existing approaches, thereby confirming its capability to improve the optimization performance and efficiency in solving practical TOPs.