Deep-Learning-Based Evolutionary Operators

Deep-learning-based genetic operators can substantially improve evolutionary search quality, but they also introduce significant computational and energy overhead, bringing the Red AI vs. Green AI tension into evolutionary computation. In this work, we study the energy footprint of Deep Neural Crossover, a reinforcement-learning-based multi-parent crossover operator, and evaluate parameter-level strategies to reduce its power consumption without sacrificing solution quality. We profile experiments with process-level energy measurements and analyze two optimization levers: training batch size and a new fitness-based scheduling threshold that triggers backpropagation only after meaningful fitness improvement. Experiments on benchmark Bin-Packing and Graph Coloring instances show that changing batch size alone has little effect on total energy, whereas the proposed scheduling strategy yields substantial savings. In particular, using larger scheduling thresholds considerably reduces energy consumption while keeping fitness close to the best-performing configurations and, in several cases, matching or improving it. These results provide a practical pathway for greener neuro-evolutionary algorithms through adaptive training schedules.

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.

We present a novel multi-parent crossover operator in genetic algorithms (GAs) called "Deep Neural Crossover" (DNC). Unlike conventional GA crossover operators that rely on a random selection of parental genes, DNC leverages the capabilities of deep reinforcement learning (DRL) and an encoder-decoder architecture to select the genes. Specifically, we use DRL to learn a policy for selecting promising genes. The policy is stochastic, to maintain the stochastic nature of GAs, representing a distribution for selecting genes with a higher probability of improving fitness. Our architecture features a recurrent neural network (RNN) to encode the parental genomes into latent memory states, and a decoder RNN that utilizes an attention-based pointing mechanism to generate a distribution over the next selected gene in the offspring. The operator’s architecture is designed to find linear and nonlinear correlations between genes and translate them to gene selection. To reduce computational cost, we present a transfer-learning approach, wherein the architecture is initially trained on a single problem within a specific domain and then applied to solving other problems of the same domain. We compare DNC to known operators from the literature over two benchmark domains, outperforming all baselines.

Traditional genetic algorithm mutation strategies typically apply stochastic gene modifications without considering broader genomic context or learning from historical evolutionary data. To address this, we introduce BERT Mutation for Genetic Algorithms (GAs), a novel, domain-independent contextual mutation operator that adapts the Transformer architecture to fixed-length genetic representations. While deep learning-based operators have shown success in Genetic Programming, extending them to GAs is challenged by “semantic opacity”—the lack of explicit structural definitions in linear integer or binary vectors. To overcome this, we propose a novel data augmentation strategy that generates pseudo-training examples derived from the population’s elite. This strategy enriches the training signal by constructing augmented trajectories toward high-fitness solutions, stabilizing the reinforcement learning process. We evaluate the operator across four diverse domains: Frozen Lake, Artificial Ant, Graph Coloring, and Unweighted Set Cover. Our results demonstrate that BERT Mutation consistently outperforms traditional stochastic operators (One-Point, Uniform, and Zigzag) and an adaptive operator (AOS) in both convergence speed and final solution quality. Notably, BERT Mutation achieves superior average fitness in approximately half the wall-clock time and fewer than half the generations required by competing baselines. Furthermore, diversity analysis confirms that BERT Mutation preserves structured, meaningful population diversity, avoiding the premature convergence seen in simpler operators.

We introduce BERT mutation, a novel, domain-independent mutation operator for Genetic Programming (GP) that leverages advanced Natural Language Processing (NLP) techniques to improve convergence, particularly using the Masked Language Modeling approach. By combining the capabilities of deep reinforcement learning and the BERT transformer architecture, BERT mutation intelligently suggests node replacements within GP trees to enhance their fitness. Unlike traditional stochastic mutation methods, BERT mutation adapts dynamically by using historical fitness data to optimize mutation decisions, resulting in more effective evolutionary improvements. Through comprehensive evaluations across three benchmark domains, we demonstrate that BERT mutation significantly outperforms conventional and state-of-the-art mutation operators in terms of convergence speed and solution quality. This work represents a pivotal step toward integrating state-of-the-art deep learning into evolutionary algorithms, pushing the boundaries of adaptive optimization in GP.

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.

In Genetic Algorithms (GAs), the selection operator plays a critical role in balancing exploration and exploitation. However, classical and adaptive selection mechanisms largely rely on static rules or handcrafted heuristics that fail to adapt to the real-time dynamics of the evolving population. In this work, we introduce Deep Tournament Selection (DTS), a novel domain-independent selection operator that reformulates tournament selection as a Markov Decision Process optimized via reinforcement learning. DTS evaluates candidate solutions in a tournament using a Transformer encoder augmented with global and local rank-based positional encodings, along with a self-attention pointer mechanism. This architecture allows the selection policy to jointly capture global fitness relationships and local competitive context. The policy is trained fully online using policy-gradient reinforcement learning without requiring additional fitness evaluations, enabling the operator to dynamically adjust its selection pressure. We evaluate DTS on three canonical combinatorial optimization domains: Graph Coloring, Set Cover, and the Traveling Salesman Problem. Empirical results demonstrate that DTS achieves faster convergence, improved solution quality, and robust performance across problem instances compared to classical and dynamic selection baselines, while introducing negligible computational overhead and preserving population diversity.

We present a novel approach to performing fitness approximation in genetic algorithms (GAs) using machine learning (ML) models, focusing on dynamic adaptation to the evolutionary state. We compare different methods for (1) switching between actual and approximate fitness, (2) sampling the population, and (3) weighting the samples. Experimental findings demonstrate significant improvement in evolutionary runtimes, with fitness scores that are either identical or slightly lower than those of the fully run GA—depending on the ratio of approximate-to-actual-fitness computation. Although we focus on evolutionary agents in Gymnasium (game) simulators—where fitness computation is costly—our approach is generic and can be easily applied to many different domains.