# Evolutionary Learning Theory: Can Evolution Produce Learning?

**Research Hub Extension for "The Bridge"**

This document compiles foundational and recent research on whether evolutionary dynamics can produce within-lifetime learning capabilities. The central question: Can the blind process of natural selection give rise to systems that learn, adapt, and abstract during their lifetime—not just across generations?

---

## Core Framework Papers

### 1. Watson & Szathmary (2016) - "How Can Evolution Learn?"

**Citation:** Watson, R. A., & Szathmary, E. (2016). How Can Evolution Learn? *Trends in Ecology & Evolution*, 31(2), 147-157.

**URLs:**
- [PubMed](https://pubmed.ncbi.nlm.nih.gov/26705684/)
- [ScienceDirect](https://www.sciencedirect.com/science/article/abs/pii/S0169534715002931)
- [ResearchGate PDF](https://www.researchgate.net/publication/288324101_How_Can_Evolution_Learn)

**Summary:** The theory of evolution links random variation and selection to incremental adaptation. Learning theory links incremental adaptation to intelligent behavior. Watson and Szathmary argue these are not mere analogies—formal equivalences exist between learning and evolution in several scenarios: selection in sexual populations with Bayesian learning, evolution of genotype-phenotype maps with correlation learning, evolving gene regulation networks with neural network learning, and evolution of ecological relationships with distributed memory models.

**Key insight:** Evolution may possess intrinsic "learning" capabilities through structural mechanisms like recombination and developmental plasticity that allow it to "internalize" environmental regularities—much like a neural network internalizes training data.

**Connection to central question:** This paper provides the theoretical foundation for understanding how evolution itself may be a form of learning algorithm. If evolution can "learn" across generations, the question becomes whether it can bootstrap within-lifetime learning.

**Follow-up papers:**
- Blute, M. (2016). Evolution and Learning: A Response to Watson and Szathmary. *TREE*.
- Livnat, A. & Papadimitriou, C. (2016). Evolution and Learning: Used Together, Fused Together. *TREE*.
- Watson, R. A. & Szathmary, E. (2016). How Can Evolution Learn? - A Reply to Responses. *TREE*.

---

### 2. Hinton & Nowlan (1987) - "How Learning Can Guide Evolution"

**Citation:** Hinton, G. E., & Nowlan, S. J. (1987). How Learning Can Guide Evolution. *Complex Systems*, 1, 495-502.

**URLs:**
- [University of Toronto](https://www.cs.toronto.edu/~hinton/absps/evolution.htm)
- [Semantic Scholar PDF](https://www.semanticscholar.org/paper/How-Learning-Can-Guide-Evolution-Hinton-Nowlan/f9197ff9fdabd2b78bfe0602365011c6699b0d66)

**Summary:** The assumption that acquired characteristics are not inherited is often taken to imply that lifetime learning cannot guide evolution. Hinton and Nowlan demonstrate computationally that this inference is incorrect. Learning alters the shape of the fitness landscape in which evolution operates, providing evolutionary paths toward sets of co-adapted alleles. Learning organisms evolve much faster than non-learning equivalents, even though learned characteristics are not directly inherited.

**Key insight:** Learning "smooths" the fitness landscape, allowing evolution to find complex adaptations that would be inaccessible through random mutation alone. The Baldwin effect creates a bridge between phenotypic plasticity and genetic fixation.

**Connection to central question:** This is the foundational computational demonstration that learning and evolution are not independent—they interact synergistically. Evolution can produce learning because learning makes evolution more effective.

---

### 3. Maynard Smith & Szathmary (1995) - "The Major Transitions in Evolution"

**Citation:** Maynard Smith, J., & Szathmary, E. (1995). *The Major Transitions in Evolution*. Oxford University Press.

**URLs:**
- [Oxford University Press](https://global.oup.com/academic/product/the-major-transitions-in-evolution-9780198502944)
- [Nature Review](https://www.nature.com/articles/374227a0)
- [Santa Fe Institute PDF](https://wiki.santafe.edu/images/0/0e/Szathmary.MaynardSmith_1995_Nature.pdf)

**Summary:** Identifies eight major transitions in evolution—from replicating molecules to eukaryotic cells, from asexual clones to sexual populations, from solitary individuals to societies, from primate societies to human language. Each transition involves a change in how information is stored and transmitted, with smaller entities combining to form larger wholes that then become the new units of selection.

**Key insight:** Evolutionary transitions involve recursive embedding—new levels of organization emerge that subsume and constrain lower levels. The emergence of language represents the transition where cultural learning becomes a new inheritance system, potentially as important as genetic inheritance.

**Connection to central question:** The major transitions framework suggests that learning (specifically human language and culture) is itself a major evolutionary transition—a qualitative shift in how information propagates. This suggests evolution can indeed produce learning, but it may require specific structural conditions.

**Update:** Szathmary (2015). *Toward major evolutionary transitions theory 2.0*. PNAS. [Link](https://www.pnas.org/doi/10.1073/pnas.1421398112)

---

## The Baldwin Effect & Genetic Assimilation

### 4. Waddington (1942) - "Canalization of Development"

**Citation:** Waddington, C. H. (1942). Canalization of development and the inheritance of acquired characters. *Nature*, 150, 563-565.

**URLs:**
- [Wikipedia - Genetic Assimilation](https://en.wikipedia.org/wiki/Genetic_assimilation)
- [PNAS Theoretical Perspective 2023](https://www.pnas.org/doi/10.1073/pnas.2309760120)
- [Wikipedia - Canalization](https://en.wikipedia.org/wiki/Canalisation_(genetics))

**Summary:** Waddington introduced "canalization"—the ability of an organism to produce the same phenotype despite genetic or environmental variation. In his famous experiments, he exposed fruit fly embryos to ether, producing a second thorax (bithorax phenotype). After 20 generations of selection, some flies developed bithorax *without* ether treatment—the environmentally-induced trait had become genetically fixed.

**Key insight:** Phenotypic plasticity and genetic evolution are not separate—plasticity can lead evolution by exploring phenotype space, with genes later "catching up" to stabilize successful variants. The phenotype can precede and guide the genotype.

**Connection to central question:** This demonstrates a mechanism by which learned/plastic behaviors can become innate—exactly the process needed for evolution to produce learning capabilities. What an organism learns to do, its descendants may be born knowing.

---

### 5. Crispo (2007) - "The Baldwin Effect and Genetic Assimilation: Revisiting Two Mechanisms"

**Citation:** Crispo, E. (2007). The Baldwin Effect and Genetic Assimilation: Revisiting Two Mechanisms of Evolutionary Change Mediated by Phenotypic Plasticity. *Evolution*, 61(11), 2469-2479.

**URLs:**
- [Oxford Academic](https://academic.oup.com/evolut/article/61/11/2469/6853881)
- [Wiley Online](https://onlinelibrary.wiley.com/doi/10.1111/j.1558-5646.2007.00203.x)
- [PubMed](https://pubmed.ncbi.nlm.nih.gov/17714500/)

**Summary:** Clarifies the distinction between Baldwin's organic selection (plasticity influences survival, guiding future genetic evolution) and Waddington's genetic assimilation (plastic traits become canalized through selection on the developmental system). Introduces "genetic accommodation" as the modern umbrella term for heritable changes in response to novel environmental induction.

**Key insight:** There are multiple distinct mechanisms by which plasticity mediates evolutionary change. Understanding these distinctions is crucial for predicting when and how learning might become genetically encoded.

**Connection to central question:** Provides the conceptual vocabulary for understanding how within-lifetime learning could become across-generation inheritance—the key transition required for evolution to "bootstrap" learning capabilities.

---

## Evo-Devo and Phenotypic Plasticity

### 6. West-Eberhard (2003) - "Developmental Plasticity and Evolution"

**Citation:** West-Eberhard, M. J. (2003). *Developmental Plasticity and Evolution*. Oxford University Press.

**URLs:**
- [Amazon](https://www.amazon.com/Developmental-Plasticity-Evolution-Mary-West-Eberhard/dp/0195122356)
- [Wikipedia - Evo-Devo](https://en.wikipedia.org/wiki/Evolutionary_developmental_biology)

**Summary:** A groundbreaking 800-page synthesis arguing that genes are often followers, not leaders, of the phenotype. West-Eberhard demonstrates through exhaustive empirical review that organisms are universally responsive to environment—plasticity is the norm, not the exception. Environmental induction may be more important than random mutation in generating the variation that evolution acts upon.

**Key insight:** "Genes are followers, not leaders, in evolutionary change." Development is inherently plastic, and this plasticity is the engine of evolutionary novelty. The organism is not merely a readout of genetic information but an active constructor of its phenotype.

**Connection to central question:** If phenotypic plasticity (a form of within-lifetime adaptation) is the primary source of evolutionary novelty, then evolution inherently depends on—and thus produces—learning-like capabilities. The question shifts from "can evolution produce learning?" to "learning is already part of how evolution works."

---

### 7. Laland et al. (2015) - "The Extended Evolutionary Synthesis"

**Citation:** Laland, K. N., et al. (2015). The extended evolutionary synthesis: its structure, assumptions and predictions. *Proceedings of the Royal Society B*, 282(1813), 20151019.

**URLs:**
- [PMC Full Text](https://pmc.ncbi.nlm.nih.gov/articles/PMC4632619/)
- [Royal Society](https://royalsocietypublishing.org/doi/10.1098/rspb.2015.1019)
- [PubMed](https://pubmed.ncbi.nlm.nih.gov/26246559/)

**Summary:** Outlines the Extended Evolutionary Synthesis (EES), which incorporates developmental bias, inclusive inheritance (genetic, epigenetic, behavioral, symbolic), and niche construction as evolutionary causes alongside natural selection. The EES predicts that phenotypic change can precede genotypic change ("plasticity-first evolution").

**Key insight:** The EES framework explicitly recognizes multiple inheritance systems—not just genes. Behavioral and cultural inheritance are legitimate evolutionary mechanisms. This means learning can evolve not just as a trait but as an inheritance channel.

**Connection to central question:** The EES provides theoretical justification for viewing learning as an evolutionary mechanism, not just an evolutionary product. Evolution produces learning because learning *is* evolution operating through a different substrate.

---

## Gene-Culture Coevolution

### 8. Boyd & Richerson (1985/2010) - "Culture and the Evolutionary Process"

**Citation:** Boyd, R., & Richerson, P. J. (1985). *Culture and the Evolutionary Process*. University of Chicago Press.

**Updated review:** Richerson, P. J., Boyd, R., & Henrich, J. (2010). Gene-culture coevolution in the age of genomics. *PNAS*, 107(S2), 8985-8992.

**URLs:**
- [Amazon - Original Book](https://www.amazon.com/Culture-Evolutionary-Process-Robert-Boyd/dp/0226069338)
- [PNAS 2010](https://www.pnas.org/doi/10.1073/pnas.0914631107)
- [NCBI Book Chapter](https://www.ncbi.nlm.nih.gov/books/NBK210012/)

**Summary:** Dual inheritance theory models humans as subject to two interacting evolutionary systems: genetic and cultural. Culture evolves through Darwinian processes (variation, selection, transmission), often faster than genes. Cultural innovations create novel selective environments for genes. The interaction produces phenomena impossible under either system alone—large-scale cooperation, cumulative technology, moral systems.

**Key insight:** Culture is an evolved inheritance system that itself evolves. Gene-culture coevolution can explain uniquely human traits like cooperation among strangers. Lactose tolerance is a classic example: dairy farming (cultural) created selection pressure for lactase persistence (genetic).

**Connection to central question:** Gene-culture coevolution demonstrates that evolution has already produced a learning system (culture) that now shapes genetic evolution. This is existence proof that evolution can produce learning—we are living in the result.

---

## Neuroevolution and Machine Learning

### 9. Stanley & Miikkulainen (2002) - "NEAT: Evolving Neural Networks Through Augmenting Topologies"

**Citation:** Stanley, K. O., & Miikkulainen, R. (2002). Evolving neural networks through augmenting topologies. *Evolutionary Computation*, 10(2), 99-127.

**URLs:**
- [MIT Press](https://direct.mit.edu/evco/article/10/2/99/1123/Evolving-Neural-Networks-through-Augmenting)
- [PubMed](https://pubmed.ncbi.nlm.nih.gov/12180173/)
- [Wikipedia](https://en.wikipedia.org/wiki/Neuroevolution_of_augmenting_topologies)

**Summary:** NEAT is a genetic algorithm for evolving both the topology and weights of neural networks. Key innovations: (1) historical markings for crossover between different topologies, (2) speciation to protect structural innovation, (3) starting minimal and complexifying. Won "Outstanding Paper of the Decade" award.

**Key insight:** Evolution can discover neural network architectures, not just optimize weights. The algorithm both optimizes and complexifies, strengthening the analogy with biological evolution. Minimal starting points lead to efficient solutions.

**Connection to central question:** NEAT demonstrates that artificial evolution can produce learning systems (neural networks). The question becomes: can this be extended to produce networks that themselves learn during deployment?

---

### 10. Stanley et al. (2019) - "Designing Neural Networks Through Neuroevolution"

**Citation:** Stanley, K. O., Clune, J., Lehman, J., & Miikkulainen, R. (2019). Designing neural networks through neuroevolution. *Nature Machine Intelligence*, 1(1), 24-35.

**URLs:**
- [Nature Machine Intelligence](https://www.nature.com/articles/s42256-018-0006-z)
- [ResearchGate PDF](https://www.researchgate.net/publication/330203191_Designing_neural_networks_through_neuroevolution)

**Summary:** Comprehensive review of neuroevolution showing it enables capabilities unavailable to gradient-based approaches: learning activation functions, hyperparameters, architectures, and even learning algorithms themselves. Neuroevolution maintains population diversity, enabling extreme exploration and massive parallelization.

**Key insight:** Neuroevolution can discover "the algorithms for learning themselves"—not just network weights or architectures, but the learning rules that govern weight updates. This is meta-learning through evolution.

**Connection to central question:** This directly addresses how evolution might produce learning: by evolving the learning algorithms themselves. The outer loop (evolution) produces the inner loop (learning).

---

### 11. Real et al. (2020) - "AutoML-Zero: Evolving Machine Learning Algorithms From Scratch"

**Citation:** Real, E., Liang, C., So, D. R., & Le, Q. V. (2020). AutoML-Zero: Evolving Machine Learning Algorithms From Scratch. *ICML 2020*.

**URLs:**
- [arXiv](https://arxiv.org/abs/2003.03384)
- [Google Research Blog](https://research.google/blog/automl-zero-evolving-code-that-learns/)
- [GitHub](https://github.com/google-research/google-research/tree/master/automl_zero)

**Summary:** AutoML-Zero evolves complete machine learning algorithms using only basic mathematical operations (addition, multiplication, variable assignment). Starting from empty programs, evolution discovers neural networks trained by backpropagation. When optimized on CIFAR-10, evolved algorithms include bilinear interactions, normalized gradients, weight averaging, and dropout-like mechanisms.

**Key insight:** Evolution can discover from scratch what took humans decades to develop. The search space is astronomically large (~10^12 candidates for an accurate algorithm), yet evolutionary search succeeds. This demonstrates that learning algorithms are evolvable.

**Connection to central question:** AutoML-Zero is direct evidence that evolution can produce learning algorithms. The limitation is that it requires a fitness function based on learning performance—the meta-learning objective is engineered, not emergent.

---

### 12. Clune (2019) - "AI-Generating Algorithms"

**Citation:** Clune, J. (2019). AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence. *arXiv preprint* arXiv:1905.10985.

**URLs:**
- [arXiv](https://arxiv.org/abs/1905.10985)
- [Apple Podcast Interview](https://podcasts.apple.com/us/podcast/accelerating-intelligence-with-ai-generating/id1116303051?i=1000588923632)
- [LessWrong Summary](https://www.lesswrong.com/posts/CfHahoLCykJuagdcj/ai-gas-ai-generating-algorithms-an-alternate-paradigm-for)

**Summary:** Proposes AI-Generating Algorithms (AI-GAs) as an alternative to manual AI engineering. Three pillars: (1) meta-learning architectures, (2) meta-learning learning algorithms, (3) generating effective learning environments. The key insight is that Darwinian evolution produced human intelligence through a simple underlying process—perhaps we can recreate this computationally.

**Key insight:** The path to general AI may be through systems that automatically learn how to learn. Open-ended evolution—processes that continually generate novelty without plateau—may be essential.

**Connection to central question:** Clune argues that evolution *did* produce learning (in biological organisms), and we should replicate this process computationally. The three pillars provide a research agenda for how.

---

## Computational Evolution of Learning

### 13. Livnat & Papadimitriou (2016) - "Sex as an Algorithm"

**Citation:** Livnat, A., & Papadimitriou, C. (2016). Sex as an algorithm: the theory of evolution under the lens of computation. *Communications of the ACM*, 59(11), 84-93.

**URLs:**
- [CACM](https://dl.acm.org/doi/10.1145/2976749)
- [PNAS: Algorithms, Games, and Evolution](https://www.pnas.org/doi/10.1073/pnas.1406556111)
- [Berkeley News](https://news.berkeley.edu/2014/06/16/algorithm-explains-sex-in-evolution/)

**Summary:** Demonstrates formal equivalence between sexual evolution under weak selection and the Multiplicative Weights Update Algorithm (MWUA) from game theory and online learning. Evolution with sex can be viewed as a coordination game between genes played according to a powerful learning algorithm.

**Key insight:** Evolution is literally a learning algorithm, not merely analogous to one. The equations of population genetics under sexual selection are identical to equations for online learning. Evolution "hedges its bets" like a gambler using optimal strategy.

**Connection to central question:** This provides mathematical proof that evolution implements learning. The question becomes: at what level? Population-level learning across generations is established—can this scaffold within-organism learning?

---

### 14. Wagner & Altenberg (1996) - "Complex Adaptations and the Evolution of Evolvability"

**Citation:** Wagner, G. P., & Altenberg, L. (1996). Complex adaptations and the evolution of evolvability. *Evolution*, 50(3), 967-976.

**URLs:**
- [Swarthmore PDF](https://www.sccs.swarthmore.edu/users/08/bblonder/phys120/docs/wagner.pdf)
- [PLOS Computational Biology - Modularity](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1000719)

**Summary:** Introduces the concept of "evolvability" as an evolved trait—the capacity to generate heritable phenotypic variation. Modularity (few pleiotropic effects between functional units) increases evolvability by allowing independent optimization of different functions.

**Key insight:** Evolution evolves its own evolvability. Modular genotype-phenotype maps are more evolvable, and modularity itself evolves. This is a form of meta-evolution—evolution optimizing its own search process.

**Connection to central question:** If evolution can evolve its own evolvability, it demonstrates a capacity for self-improvement that parallels learning. The architecture that enables evolution to learn may have evolved.

---

## Meta-Learning and the Inner/Outer Loop

### 15. Finn et al. (2017) - "Model-Agnostic Meta-Learning" (MAML)

**Citation:** Finn, C., Abbeel, P., & Levine, S. (2017). Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. *ICML 2017*.

**URLs:**
- [arXiv](https://arxiv.org/abs/1703.03400)
- [BayesWatch Tutorial](https://www.bayeswatch.com/2018/11/30/HTYM/)

**Summary:** MAML trains a neural network initialization such that a few gradient steps on a new task produce good performance. The outer loop optimizes the initialization; the inner loop adapts to specific tasks. This mirrors evolution (outer) producing learning (inner).

**Key insight:** The inner/outer loop structure of meta-learning maps directly to the evolution/learning structure in biology. The outer loop (meta-training) is like evolution—slow optimization across many tasks. The inner loop (adaptation) is like learning—fast optimization within a task.

**Connection to central question:** MAML demonstrates how an outer optimization process can produce effective inner optimization. The question is whether evolution as an outer loop can produce this structure without explicit meta-learning objectives.

---

## Summary Table

| Source | Year | Core Claim | Evidence Type |
|--------|------|------------|---------------|
| Watson & Szathmary | 2016 | Formal equivalences between evolution and learning | Theoretical |
| Hinton & Nowlan | 1987 | Learning smooths fitness landscape for evolution | Computational |
| Maynard Smith & Szathmary | 1995 | Major transitions involve new inheritance systems | Theoretical/Empirical |
| Waddington | 1942 | Plastic traits can become genetically fixed | Experimental |
| Crispo | 2007 | Multiple mechanisms mediate plasticity-evolution interaction | Review |
| West-Eberhard | 2003 | Genes follow phenotypes; plasticity drives novelty | Synthesis |
| Laland et al. | 2015 | Extended synthesis: multiple inheritance channels | Framework |
| Boyd & Richerson | 1985/2010 | Gene-culture coevolution produces learning | Theoretical/Empirical |
| Stanley & Miikkulainen | 2002 | NEAT evolves neural network topologies | Computational |
| Stanley et al. | 2019 | Neuroevolution can discover learning algorithms | Review |
| Real et al. | 2020 | AutoML-Zero evolves ML algorithms from scratch | Computational |
| Clune | 2019 | AI-GAs: three pillars for evolving intelligence | Theoretical/Programmatic |
| Livnat & Papadimitriou | 2016 | Sexual evolution = MWUA learning algorithm | Mathematical |
| Wagner & Altenberg | 1996 | Evolution evolves its own evolvability | Theoretical |

---

## Synthesis: What Do We Know?

### Established

1. **Evolution is formally equivalent to learning** at the population level (Livnat & Papadimitriou).
2. **Learning accelerates evolution** through the Baldwin effect (Hinton & Nowlan, Waddington).
3. **Plastic traits can become genetic** through genetic assimilation (Waddington, Crispo).
4. **Evolution has produced cultural learning** as a new inheritance system (Boyd & Richerson).
5. **Artificial evolution can discover learning algorithms** (AutoML-Zero, NEAT).

### Open Questions

1. Can evolution produce within-lifetime learning **without explicit meta-learning objectives**?
2. What environmental conditions are necessary and sufficient?
3. Is there a minimal complexity threshold for learning emergence?
4. Can the transition from plastic behavior to genetic encoding scale to abstract reasoning?

### Implications for Bridge Research

The literature suggests that the gap between replication and reasoning may not be as large as it appears. Evolution and learning are deeply intertwined—perhaps different expressions of the same underlying process of adaptation. The research program should focus on:

1. **Creating environmental conditions** that make within-lifetime adaptation advantageous
2. **Measuring the transition** from plastic behavior to stable internal models
3. **Testing for transfer** to novel tasks not present during evolution
4. **Avoiding engineered meta-objectives** to test genuine emergence

---

## Further Reading Directions

- **Hopfield Networks**: Connection between associative memory and evolutionary dynamics
- **Free Energy Principle**: Friston's framework unifying perception, action, and learning
- **Open-Ended Evolution**: Conditions for sustained novelty generation
- **Developmental Bias**: How development constrains and enables evolutionary change
- **Predictive Processing**: Learning as prediction error minimization

---

*Last updated: 2026-02-17*
*Compiled for the Virtue Research Hub - Bridge Section*