Adaptive Dimensional Learning for Multitask Optimization

Evolutionary multitask optimization (EMTO) is an emerging optimization approach that leverages knowledge transfer (KT) across multiple tasks to accelerate convergence and enhance solution quality. However, existing EMTO methods typically focus predominantly on index-aligned learning during self-evolution and KT, overlooking the potential of index-unaligned dimensions. Since the index-aligned dimensions differ in their representation or significance across tasks, strictly index-aligned learning may not be suitable, because it may lead to ineffective or even negative transfer. To address this and fully harness the potential of index-unaligned learning, this paper proposes an adaptive dimensional learning (ADL) for multitask optimization (ADLMTO), which utilizes both index-aligned and index-unaligned dimensional information to enhance optimization effectiveness by implementing evolutionary learning at the dimensional level. ADLMTO employs the adaptive feedback-driven evolutionary strategy (AFDES) to dynamically adjust evolutionary modes, including index-aligned and index-unaligned learning during both self-evolution and KT. ADL strategy performs effective dimensional-level learning based on evolutionary modes, thereby facilitating enhanced optimization. Experimental results demonstrate that ADLMTO outperforms state-of-the-art EMTO algorithms on the CEC2017 and CEC2022 multitask evolution benchmarks, confirming the effectiveness of ADLMTO.