<p>Solving constrained multi-objective optimization problems (CMOPs) is a popular research direction in the field of optimization. To address the limitations of using a single strategy in the recent constrained multi-objective evolutionary algorithms (CMOEAs), reinforcement learning (RL) techniques have been applied to solve CMOPs and have achieved remarkable results. However, existing methods use a relatively simplistic state representation to train RL model, which impairs the decision-making capability due to the insufficient training. Moreover, the correlations among tasks are not considered when selecting the strategy, potentially leading to negative knowledge transfer. In this paper, a CMOEA based on attention mechanism assisted auxiliary task selection (AMTCMO) is proposed. Firstly, a population feature fusion method based on self-attention mechanism is used to generate the real-time state of the population for training the RL model. Then, the enhanced state can provide dynamic environment perception ability for the RL model, which can accurately reflect the auxiliary task needed at present. Secondly, a selection strategy based on cross-attention mechanism driven by task association is proposed, which can analyze the benefit of each task. The obtained attention scores and RL collaboratively decide the auxiliary task selection to enhance the reliability. Experimental results show that the proposed AMTCMO demonstrates excellent performance on 4 test suites and 2 real-world engineering problems compared with nine state-of-the-art CMOEAs.</p>