A dual prediction strategy with inverse model for evolutionary dynamic multiobjective optimization.

Area of Science:

• Optimization
• Evolutionary Computation
• Artificial Intelligence

Background:

• Dynamic multiobjective optimization problems (DMOPs) are common in real-world applications.
• Existing prediction-based strategies for multiobjective evolutionary algorithms (MOEAs) face challenges in maintaining prediction accuracy.
• Tracking moving Pareto Fronts and finding Pareto Sets in dynamic environments is crucial.

Purpose of the Study:

• To develop a robust prediction strategy for DMOPs that mitigates the impact of inaccurate predictions.
• To enhance the performance of MOEAs in dynamic environments by improving prediction accuracy and search guidance.
• To address the challenge of mapping solutions between decision and objective spaces effectively.

Main Methods:

• A dual prediction strategy with an inverse model (DPIM) was developed.
• DPIM predicts individuals in the objective space upon detecting environmental changes.
• An inverse model was established to map between decision and objective spaces, with predictions to minimize mapping errors.

Main Results:

• The effectiveness of DPIM was validated against four state-of-the-art dynamic multiobjective evolutionary algorithms (DMOEAs).
• Experiments were conducted on 14 benchmark problems simulating various real-world scenarios.
• DPIM demonstrated its ability to achieve high-quality populations with superior convergence and distribution.

Conclusions:

• The proposed DPIM effectively alleviates the negative impact of inaccurate predictions in dynamic multiobjective optimization.
• DPIM provides a reliable method for guiding the search in dynamic environments, leading to improved solution quality.
• The strategy shows significant promise for practical applications involving DMOPs.