Deep Learning-Based Operators for Evolutionary Algorithms (2407.10477v1)

Abstract: We present two novel domain-independent genetic operators that harness the capabilities of deep learning: a crossover operator for genetic algorithms and a mutation operator for genetic programming. Deep Neural Crossover leverages the capabilities of deep reinforcement learning and an encoder-decoder architecture to select offspring genes. BERT mutation masks multiple gp-tree nodes and then tries to replace these masks with nodes that will most likely improve the individual's fitness. We show the efficacy of both operators through experimentation.

• The paper proposes the Deep Neural Crossover (DNC) operator that employs deep reinforcement learning to capture complex gene correlations for enhanced GA performance.
• The paper introduces BERT Mutation, which adapts masked language modeling to optimize GP mutation, achieving superior fitness on benchmark tests.
• The paper demonstrates that applying transfer-learning reduces DRL computational overhead, broadening applicability across scheduling and combinatorial optimization tasks.

Deep Learning-Based Operators for Evolutionary Algorithms

Overview

The paper "Deep Learning-Based Operators for Evolutionary Algorithms" by Eliad Shem-Tov, Moshe Sipper, and Achiya Elyasaf introduces two novel deep learning-driven operators designed to improve the efficacy of genetic algorithms (GAs) and genetic programming (GP). Specifically, the paper presents a multi-parent crossover operator dubbed Deep Neural Crossover (DNC) and a mutation operator based on the architecture of BERT, named BERT Mutation.

Deep Neural Crossover (DNC)

The DNC operator leverages deep reinforcement learning (DRL) coupled with an encoder-decoder architecture to dynamically select advantageous genes during the crossover process in GAs. Unlike traditional crossover methods such as one-point or uniform crossover, which rely on random selection, DNC uses an RNN-based model to encode parental genomes and a pointing mechanism to decode and generate offspring by learning gene correlations. This approach eliminates biases inherent in traditional methods and allows for capturing both linear and nonlinear relationships between genes.

The paper highlights the computational intensity of training a DRL model but proposes a transfer-learning strategy. This involves initially training the DNC on a specific problem within a domain and subsequently applying the trained model to other problems in the same domain, thus reducing the computational overhead.

In evaluation, DNC outperformed all traditional crossover methods in terms of solution quality over benchmark domains such as Graph Coloring and Bin Packing, showing significant improvements with notably better fitness scores and convergence rates.

BERT Mutation

BERT Mutation adapts principles from NLP, particularly the Masked LLMing (MLM) of BERT to the GP mutation process. This operator masks multiple nodes in a GP tree and uses a trained BERT model to predict replacements that are likely to enhance individual fitness. Instead of relying on ground-truth-based training, BERT Mutation employs reinforcement learning, optimizing for fitness improvement as a reward. The method takes into account the structural context of the GP tree, ensuring valid replacements through a constrained sampling process.

The experimental results demonstrate that BERT Mutation significantly outperforms traditional GP mutation operators such as point mutation, subtree mutation, and hoist mutation across various benchmark datasets, including both standard regression problems and specially designed symbolic regression tasks. Notably, BERT Mutation achieves superior fitness while maintaining competitive runtime performance.

Implications and Future Directions

The practical implications of these findings are manifold. By utilizing deep learning to inform crossover and mutation processes, the evolutionary algorithms can explore the solution space more effectively, potentially leading to better optimization outcomes in various domains including scheduling, feature selection, and combinatorial problems. The transfer-learning concept applied in DNC showcases an efficient way to mitigate the computational load typically associated with DRL.

Theoretically, these findings suggest a compelling synergy between evolutionary algorithms and deep learning techniques, opening up avenues for further research. Possible future directions include expanding the DNC approach to additional domains, exploring the state maintenance capabilities of BERT Mutation more extensively in dynamic environments, and integrating these operators into hybrid evolutionary frameworks.

Conclusion

In conclusion, the methods proposed by Shem-Tov et al. mark a significant step forward in enhancing the operability and effectiveness of evolutionary algorithms through deep learning techniques. By addressing traditional shortcomings in crossover and mutation operators, these innovations set a new benchmark for domain-independent genetic operators, promising broad applicability and substantial performance gains across various evolutionary computing tasks.