Dynamic Multi-objective Optimisation Problems (DMOOPs) challenge traditional evolutionary algorithms, as the Pareto-optimal front and set evolve over time. Conventional reactive methods suffer from an adaptation lag, responding only after environmental changes. To address this, we propose a proactive optimisation framework integrating Long Short-Term Memory (LSTM) networks with multi-objective evolutionary algorithms. Our approach leverages the LSTM's ability to capture complex nonlinear temporal dependencies to forecast future problem states. A key contribution of this work is a hybrid population management strategy, employing a 70/30 split: 70\% of the individuals are generated based on predicted future states to enhance exploitation, while the remaining 30\% are exploratory individuals designed to maintain population diversity and robustness against prediction errors. Experimental results on standard FDA and dMOP benchmarks demonstrate that our method significantly outperforms reactive baselines under computationally constrained scenarios, achieving up to a 50\% improvement in hypervolume over conventional NSGA-II, while maintaining a robust advantage over state-of-the-art prediction-based approaches like TR-DMOEA and PPS. Furthermore, validation on a data centre thermal management problem confirms that LSTM-EA successfully maintains high thermal safety margins while optimising energy costs. These findings suggest that deep-learning-enhanced proactive optimisation is highly effective for managing dynamic environments in both theoretical and industrial applications.
Masayuki Kaneko (Tue,) studied this question.