- The paper demonstrates that an RBM-based multivariate EDA outperforms traditional random mutation models in high-dimensional optimization tasks by effectively integrating gene-gene interactions.
- The model captures epistatic dependencies via Restricted Boltzmann Machines, enabling accurate representation of internal, heritable information without external signals.
- Results imply that nonrandom, non-Lamarckian mutation mechanisms can drive adaptive evolution and challenge classical evolutionary assumptions in computational models.
Nonrandom, Non-Lamarckian Mutation in Evolution: A Multivariate Estimation-of-Distribution Model
Introduction and Theoretical Context
This work proposes and explores an abstract computational model for evolutionary processes in which mutation is neither random in the classical sense nor Lamarckian. The foundational dichotomy distinguishing standard evolutionary models—random mutation under natural selection versus Lamarckian inheritance—has been dominant for a century. The Interaction-Based Evolution (IBE) framework, however, postulates a third regime where the probabilities of individual mutational events are influenced by the accumulation of internal, heritable information over generations, not directly by environmental conditions nor through undirected accident.
Empirical evidence motivating IBE includes cases such as the human hemoglobin S mutation and the APOL1 1024A→G mutation, where the de novo occurrence rates of specific beneficial mutations are elevated in those populations and loci where they confer adaptive advantage, contradicting expectations from models where mutation is random at the level of individual sites. This phenomenon is unexplained by standard modifier models, which admit only adaptation of average mutation rates over many loci, not locus-specific adaptivity.
Estimation-of-Distribution Algorithm Framework and Model Specification
The paper formalizes an evolutionary simulation model using a multivariate Estimation-of-Distribution Algorithm (EDA) grounded in Restricted Boltzmann Machines (RBMs). In contrast to genetic algorithms (GAs) utilizing recombination and random mutation, EDAs replace explicit variation operators with statistical modeling: at each generation, a statistical distribution is learned from selected individuals, and new individuals are sampled from this updated distribution.
This multivariate EDA framework instantiates nonrandom, non-Lamarckian mutation at an abstract but operational level: the distribution is updated exclusively via internal information—survivors' genomes—without any direct environmental signal or knowledge of the selection landscape. Critically, the RBM structure captures gene-gene dependencies, i.e., epistatic effects, during the estimation step. The analogy is explicitly drawn between the Hebbian learning rule ("units that fire together wire together") governing RBM weight updates and the observed mutational propensity for genes "used together" to become physically linked or fused in biological systems.
The simulation focuses on population evolution under threshold selection on fitness functions defined by combinatorial optimization instances—primarily MAX-k-SAT problems—configured to test high-dimensional, epistatic fitness landscapes. Individuals are represented as binary vectors, and fitness is determined as the proportion of maximally satisfied clauses.
Key Numerical and Qualitative Results
The simulations yield several robust and counterintuitive results:
- Superiority of Nonrandom, Non-Lamarckian Mutation in Complex Optimization: For high-dimensional MAX-3-SAT instances (n=150, $600$, $2400$), the RBM-based EDA consistently achieves higher mean population fitness than both random mutation (RM) models and pure random guess (BRG) baselines. The advantage of the RBM model increases with both the genome size and population size.
- Dependence on Epistatic Interactions: Direct tests using synthetic parity-function fitness landscapes (with multiple maxima and defined epistatic dependencies) demonstrate the RBM's ability to capture and utilize higher-order gene-gene relationships in the generation of heritable information, as shown by higher fitness in genome completion tests relative to shuffled controls.
- Role of Internal vs. External Information: Comparative ablations (weights-only vs. biases-only RBM variants) show that the RBM's fitness gains are primarily due to modeling gene-gene interactions (weights) rather than independent locus effects (biases), conclusively linking adaptive improvements to the uptake and use of internal, multivariate information.
- Dynamic Relation Between Selection Pressure and Mutational Output: As the population converges on optimal solutions, the RBM-driven system intrinsically reduces heritable variation, mirroring the empirical decline in genetic variation observed in optimizing populations. Introduction of novel selection pressures transiently increases the amount of variation produced until the population adapts.
- Bell-Shaped Trait Distributions Without Additivity: Even with strongly interacting loci (i.e., non-additive, epistatic landscapes like MAX-SAT), the distribution of population-level fitness remains approximately Gaussian over generations and the standard deviation is sustained, providing a direct counterexample to the traditional view that bell-shaped distributions in nature require independently additive genetic contributions.
Implications for Evolutionary Theory and Computational Models
The results challenge the necessity of assuming random mutation at the level of individual mutational events in evolutionary modeling. The demonstration that robust, nonrandom, non-Lamarckian mutational mechanisms can not only sustain adaptive evolution but outperform standard models on complex combinatorial landscapes undermines the “randomness/selection” orthodoxy.
The IBE-inspired RBM-EDA model exhibits several properties with broad consequences:
- Integration of Internal Information: The model provides a formal realization of evolutionary processes in which information is aggregated via distributed, networked integration rather than exclusively via population-level selection acting on stochastic variance. This reconceptualizes evolution partially in terms of population-level “learning,” drawing sharp computational parallels to unsupervised representation learning in machine learning.
- Resolution of the Sex and Recombination Paradox: By capturing the fan-in of heritable information, the framework reconciles the role of sexual recombination with the persistence of complex structure—sexual mixing does not merely disrupt combinations but enables systematic restructuring and integration, making recombination a central, not auxiliary, feature of evolution.
- Generalization Beyond Additive Genetics: The model demonstrates that complex, interacting genetic architectures can maintain adaptive variation and typical population distributions without recourse to the additive assumptions foundational to the modern synthesis.
- Algorithmic and Machine Learning Connections: The abstraction aligns the mechanics of evolution with those of generalized statistical learning algorithms. The RBM’s use of Hebbian-like mechanisms and population-level integration of survival-linked features provides a template for further development of algorithms at the interface of evolutionary computation and deep unsupervised representation learning.
Future Directions and Theoretical Outlook
Several avenues for future research are highlighted or implied:
- Incorporation of Realistic, Evolving Mutational Mechanisms: Moving beyond the abstract, centralized RBM, incorporating naturally evolving mechanisms of mutation and recombination—grounded in empirical mechanistic biology—will be essential for constructing models that bridge abstraction and realism.
- Formal Analytical Treatment of Nonrandom, Multivariate Mutation: Developing mathematical tools for analyzing convergence, complexity, and adaptability in multivariate EDA models informed by IBE is a critical next step.
- Exploration of Multi-level Learning in Evolution: Recognizing evolution as a population-level learning process, with feedback not strictly limited to individual selection but incorporating internal integration of information over generations, provides fertile territory for formal learning-theoretic analysis of biological evolution.
- Implications for Artificial Intelligence: The convergence of evolutionary theory and statistical learning, exemplified by the RBM-EDA model, suggests principled algorithms for high-dimensional, combinatorial optimization in AI. Concepts like internal representation integration, multi-level feedback, and Hebbian-style adaptation within populations could inform new generations of adaptive systems beyond current deep learning and evolutionary algorithms.
Conclusion
This work provides a rigorous demonstration that evolution does not require randomness at the mutational event level nor direct environmental feedback (Lamarckism) to generate open-ended adaptive complexity. By formalizing nonrandom, non-Lamarckian mutation in a multivariate EDA framework, drawing operational and mechanistic parallels to empirical observations and Hebbian learning, the model opens new conceptual and algorithmic directions for both evolutionary theory and computational learning. It demonstrates the logical coherence and practical potential of substituting the random mutation assumption, with direct consequences for the modeling of evolution, evolutionary computation, and algorithmic learning systems.