This paper proposes an efficient and improved variant of the success history adaptive differential evolution (SHADE), termed Adaptive SHADE with Multi-agent Participated Competition Mechanism (ASHADE-MPC). The proposed ASHADE-MPC incorporates two key components: (1) a Multi-agent Participated Competition (MPC) mechanism and (2) an adaptive mutation probability. The incorporated MPC mechanism enhances search efficiency and stability by integrating two powerful mutation strategies: DE/cur-to- \(\text {mean}(\varvec{x}^t_{idx1}, \varvec{x}^t_{idx2})\) /1 and DE/cur-to- \(\text {mean}(\varvec{x}^t_{best}, \varvec{x}^t_{idx})\) /1. Additionally, ASHADE-MPC dynamically adjusts the implementation of DE/best/1 and the multi-agent competition mechanism through a switching probability P to balance local exploitation and global exploration at different optimization stages effectively. Extensive experiments in CEC2017, CEC2020, CEC2022, and seven engineering problems against eleven state-of-the-art optimizers demonstrate the superiority and domination of ASHADE-MPC in different optimization challenges. Furthermore, we apply ASHADE-MPC to the ensemble learning domain to detect the coffee leaf disease, where three high-accuracy deep learning models are fused using an ASHADE-MPC-optimized soft voting scheme. Experimental results confirm that the proposed ASHADE-MPC-Ensemble approach improves the accuracy of 0.952% in accuracy, 0.938% in precision, 0.952% in recall, and 0.952% in F1-score compared to the second-best model, Swin Transformer, which highlights the effectiveness and applicability of ASHADE-MPC in real-world scenarios. The source code of this research can be found at https://github.com/RuiZhong961230/ASHADE-MPC.
