Transfer Learning based Dynamic Multiobjective Optimization Algorithms (1612.06093v2)

Abstract: One of the major distinguishing features of the dynamic multiobjective optimization problems (DMOPs) is the optimization objectives will change over time, thus tracking the varying Pareto-optimal front becomes a challenge. One of the promising solutions is reusing the "experiences" to construct a prediction model via statistical machine learning approaches. However most of the existing methods ignore the non-independent and identically distributed nature of data used to construct the prediction model. In this paper, we propose an algorithmic framework, called Tr-DMOEA, which integrates transfer learning and population-based evolutionary algorithm for solving the DMOPs. This approach takes the transfer learning method as a tool to help reuse the past experience for speeding up the evolutionary process, and at the same time, any population based multiobjective algorithms can benefit from this integration without any extensive modifications. To verify this, we incorporate the proposed approach into the development of three well-known algorithms, nondominated sorting genetic algorithm II (NSGA-II), multiobjective particle swarm optimization (MOPSO), and the regularity model-based multiobjective estimation of distribution algorithm (RM-MEDA), and then employ twelve benchmark functions to test these algorithms as well as compare with some chosen state-of-the-art designs. The experimental results confirm the effectiveness of the proposed method through exploiting machine learning technology.

• The paper introduces Tr-DMOEA, a framework integrating transfer learning with evolutionary algorithms to address dynamic multiobjective optimization challenges.
• It leverages historical data to build an initial population and adapts to non-iid data, significantly reducing MIGD values compared to traditional methods.
• Experimental results across twelve benchmarks demonstrate Tr-DMOEA’s superior performance and adaptability in dynamic and uncertain environments.

Insights into Transfer Learning-Based Dynamic Multiobjective Optimization Algorithms

The paper introduces Transfer Learning based Dynamic Multiobjective Optimization Algorithms (Tr-DMOEA), a framework aimed at addressing the challenges inherent in Dynamic Multiobjective Optimization Problems (DMOPs). DMOPs are characterized by shifting optimization objectives over time, necessitating the continuous tracking of the Pareto-Optimal Front (POF). Existing evolutionary algorithms (EAs) often struggle with these dynamics due to their reliance on the assumption that solutions are independently and identically distributed, which is not applicable in DMOPs. The Tr-DMOEA is designed to leverage historical data through transfer learning to improve adaptability and performance.

Algorithmic Framework

Tr-DMOEA uniquely integrates transfer learning with population-based EAs, positioning the approach to accommodate dynamic shifts in optimization problems effectively:

• Transfer Learning Integration: The framework applies transfer learning to estimate and construct an initial population by reusing past experiences. This initial population is crucial for accelerating the evolutionary process in a changing optimization landscape.
• Non-IID Hypothesis: A significant contribution of this work is addressing the non-independent and identically distributed nature of data points within the context of DMOPs. Tr-DMOEA constructs a mapping through transfer learning that allows for variations in distributions, making traditional machine learning limitations less impactful.
• Population-Based Approaches: The framework is versatile, allowing integration with various established evolutionary algorithms like NSGA-II, MOPSO, and RM-MEDA. This adaptability ensures broad utility across different optimization scenarios without necessitating major modifications.

Experimental Verification

The framework's efficacy is validated through incorporation with NSGA-II, MOPSO, and RM-MEDA, tested on twelve benchmark functions representing diverse dynamic scenarios. The experimental outcomes deliver several key insights:

• Performance Metrics: The analysis utilizes IGD (Inverted Generational Distance), its variants, and reactivity measures to robustly evaluate the algorithms. Metrics such as MIGD and DMIGD enable assessments across varying environmental configurations.
• Numerical Results: Strong numerical results are presented, showcasing Tr-DMOEA's effectiveness in optimizing DMOPs. The approach significantly reduces MIGD values, indicating closer approximation to the ideal POFs relative to the baseline algorithms.
• State-of-the-art Comparison: Compared to state-of-the-art methodologies like MBN-EDA, RND, and MOEA/D-KF, Tr-DMOEA consistently demonstrates superior or comparable performance, suggesting that the integration of transfer learning enhances the adaptability of evolutionary algorithms toward dynamic changes.

Implications and Future Prospects

The integration of transfer learning into evolutionary algorithms applied to DMOPs demonstrates a practical advance in performance, particularly in terms of speed and robustness. The algorithmic framework holds substantial promise for real-world applications where dynamic and uncertain environments are prevalent. The insights drawn here could stimulate future research into more sophisticated transfer learning methodologies and further optimize parameter settings for various classes of DMOPs.

The implications extend beyond theoretical improvements, offering potential utility in complex, real-time systems needing dynamic adaptation, such as financial markets, industrial control systems, and adaptive recommendation systems. Further exploration might delve into hybrid models, combining different machine learning paradigms with evolutionary mechanisms, fortifying the approach against the multifold challenges posed by dynamic environments.