11.7 Sophistication of self-evolution - theoretical foundation and detailed design

[!NOTE] For the latest implementation status, please refer to Functional Implementation Status (Remaining Functionality).

Date: 2026-03-05 Author: Masahiro Aoki Copyright: 2026 Moonlight Technologies Inc. Status: 📅 Planning → Implementation design phase Related documents: ADVANCED_EVOLUTION_PHASE5_6.md, EVOSPIKENET_CONCEPTS.md

overview

EvoSpikeNet's L5 self-evolution layer builds on the implementation of Phase 5 (Highly Structural Mutation) and Phase 6 (Cooperative Evolution) to achieve evolutionary capabilities at a higher level of abstraction. This document details the theoretical foundations of the four subsystems that make up 11.7 Advancing Self-Evolution.

1. Meta-Evolution

1.1 Definition and motivation

Meta-evolution is a mechanism in which the parameters, structure, and strategies of the evolutionary algorithm itself evolve over time. While normal evolution (first-order evolution) optimizes the weights and structure of a neural network, meta-evolution learns and adapts ``how to evolve'' itself.

As biological counterparts, sexual reproduction, recombination frequency, and mutation rate are subject to selective pressures over long evolutionary time scales. A high mutation rate expands the search space but destroys fitness, while a low mutation rate tends to converge to a local optimum. Meta-evolution dynamically resolves this trade-off.

1.2 Theoretical framework

1.2.1 Two-layer evolutionary model

\[ \text{Evolution}(\text{Individual}) = f\left(\theta_{\text{evo}}\right) \]

\[ \text{Meta Evolution}(\theta_{\text{evo}}) = g\left(\theta_{\text{meta}}\right) \]

where: - \(\theta_{\text{evo}}\) = evolutionary parameter set (mutation rate \(\mu\), crossover probability \(p_c\), selection pressure \(k\)) - \(\theta_{\text{meta}}\) = metaparameter set (adaptation speed, exploration/exploitation balance)

1.2.2 Adaptive mutation rate theory

Self-adaptive evolutionary strategy (Self-Adaptive ES) incorporates the mutation rate \(\sigma\) into the genome and allows it to evolve together with the individual:

\[ \sigma' = \sigma \cdot \exp\left(\tau' \cdot \mathcal{N}(0,1) + \tau \cdot \mathcal{N}_i(0,1)\right) \]

\[ x' = x + \sigma' \cdot \mathcal{N}(0, \mathbf{I}) \]

where \(\tau' = 1/\sqrt{2n}\), \(\tau = 1/\sqrt{2\sqrt{n}}\) (\(n\) = problem dimension).

1.2.3 Genomic representation of evolutionary algorithms

EvoSpikeNet extends EvoGenome's global_config to incorporate evolution parameters:

@dataclass
class MetaEvolutionConfig:
    """メタ進化パラメータ（ゲノムの一部として進化する）"""
    mutation_rate: float          # Basic mutation rate μ ∈ [0.001, 0.5]
    crossover_rate: float         # Crossover probability p_c ∈ [0.0, 1.0]
    selection_pressure: float     # Selection pressure k ∈ [1.0, 10.0]
    population_sampling: str      # "tournament" | "roulette" | "rank"
    elitism_ratio: float          # Elite preservation ratio ∈ [0.0, 0.3]
    repair_strategy: str          # "none" | "clamp" | "reflect"

    # Adaptation coefficient of the metaparameter itself
    meta_lr: float = 0.01
    meta_sigma: float = 0.1

1.2.4 Meta-fitness function

The fitness of an individual in metaevolution is evaluated by the average fitness of the offspring produced by that evolutionary strategy:

\[ F_{\text{meta}}(\theta_{\text{evo}}) = \mathbb{E}_{g \sim P(\theta_{\text{evo}})}\left[F_{\text{task}}(g)\right] \]

This selects an evolutionary strategy that "produces good offspring."

1.3 Specific mechanism in EvoSpikeNet

1.3.1 Self-adaptive mutation engine

Extend AdaptiveMutationConfig in AdvancedMutationEngine (advanced_mutations.py):

Success history-based adaptation: Update the probability of each type based on the success rate of each mutation type in the last \(N\) generations
\(p_{m,t+1} = p_{m,t} + \alpha \cdot (\text{success rate}_t - \bar{p}_t)\)
1/5 Success Rule (1/5-Rule): According to the success rule of Rechenberg (1973), if the success probability exceeds \(1/5\), increase \(\sigma\), and if it falls below, decrease it.

1.3.2 Evolution of crossover strategy

Crossover type	Gene expression	Effective scene
One-point crossover	`crossover_type=1`	Simple continuous parameter
Two-point crossover	`crossover_type=2`	Module boundary protection
Uniform crossover	`crossover_type=uniform`	When independence between genes is high
Arithmetic crossover	`crossover_type=arithmetic`	Continuous value optimization
Subgenome crossover	`crossover_type=subgenome`	Recombination of chromosome units

Cross-over types themselves are also recorded in the genome and are subject to meta-evolution.

1.3.3 Dynamic switching of selection mechanisms

Detect environmental variability (non-stationarity in the fitness landscape) and switch selection strategies:

検出：fitness_variance の急増 → 探索強化 (σ 増加、selection_pressure 低減)
検出：fitness_plateau (連続 k 世代改善なし) → 活用強化 (elitism_ratio 増加)
検出：diversity_collapse (species 数 < threshold) → 多様性回復 (mutation_rate 増加)

1.4 Theoretical guarantees and limitations

Convergence theory: From the No Free Lunch theorem (Wolpert & Macready, 1997), meta-evolution also specializes in a specific problem class. Generic meta evolution is impossible.
Computational cost: Evaluation of meta-evolution requires \(O(N_{\text{gen}})\) times more cost than object evolution.
Overfitting: Risk of metaparameters overfitting a specific task → Regular diversification is required.

2. Hierarchical Evolution

2.1 Definition and motivation

Hierarchical evolution is a mechanism in which adaptation occurs simultaneously at different time scales and abstraction levels. It combines "high-level" changes with slow evolutionary rates (modular structure, connectivity patterns) and fast "low-level" changes (synaptic weights, spike thresholds) to achieve both stability and adaptability.

Biological correspondence: Corresponds to hierarchical information processing in the cerebral cortex (visual hierarchy of V1 → V2 → V4 → IT) and differences in learning speed.

2.2 Theoretical framework

2.2.1 Time scale separation theory

Separating evolution across multiple timescales (corresponding to Maclean & Tononi, 2019's theory of consciousness):

Level	Target of change	Timescale	Implementation
L0 (reflex)	Spike threshold, immediate response	ms to s	STDP, Homeostasis
L1 (learning)	Synaptic weight, plasticity coefficient	s ~ minutes	Plasticity rules
L2 (developmental)	Network structure, layer size	Minutes to hours	AdvancedMutationEngine
L3 (Evolution)	Module connections, genome structure	Time to day	EvolutionEngine
L4 (Meta Evolution)	Evolution Strategy Parameters	Days to Weeks	MetaEvolution

2.2.2 Interlayer signaling

Higher levels set "constraints" to lower levels, and adaptation results of lower levels are "feedback" to higher levels:

\[ \text{constraint}_{L_i \to L_{i-1}} = h_i(\text{state}_{L_i}) \]

\[ \text{feedback}_{L_{i-1} \to L_i} = \nabla_{L_i} F(\text{perf}_{L_{i-1}}) \]

2.2.3 Hierarchical genome representation

The current EvoGenome has a flat Chromosome structure. Hierarchical evolution organizes chromosomes into layers:

EvoGenome
├── MetaChromosome        (L4: メタ進化戦略)
│   └── EvoStrategyGenes
├── MacroChromosome       (L3: モジュール間接続)
│   └── ConnectivityGenes
├── MesoChromosome[]      (L2: 各モジュール構造)
│   ├── LayerStructureGenes
│   └── TopologyGenes
└── MicroChromosome[]     (L1: シナプス可塑性)
    └── PlasticityGenes

2.2.4 Dynamic adjustment of evolution speed

The evolutionary rate \(v_i\) at each level is automatically adjusted by the size of the fitness gradient:

\[ v_i(t) = v_{i,\text{base}} \cdot \left(1 + \beta \cdot \left|\frac{\partial F}{\partial \theta_i}\right|\right) \]

Large fitness gradient: change in that level is effective → speed increase
Small fitness gradient: its level is saturated → speed reduction saves computational resources

2.2.5 Utilizing the Baldwin effect

Implementing the “Baldwin effect” where learning (plasticity) directs genetic evolution:

Individuals adapt to the environment through short-term learning
Use parameter values near the end point of learning as targets for “genetic assimilation”
Directing gene-level optimization to the learned parameter space

2.3 Specific mechanism in EvoSpikeNet

2.3.1 Integration with plasticity hierarchy

Integrate evolution layer based on existing hierarchical_plasticity.py:

Register each level learning rate of HierarchicalPlasticityController as a gene of EvoGenome
Feedback of performance indicators (convergence speed, stability) of plasticity rules to higher evolutionary levels

2.3.2 Evolution of module connections

Extend connection_matrix of network_topology of Chromosome between modules:

Macro connection matrix: Information flow between modules (PFC, Memory, Vision)
Micro connection matrix: Interlayer connections within a module
Optimized both levels with independent evolution speeds

2.3.3 Brain region role model

Hierarchical roles corresponding to the 24-node configuration of the distributed brain:

抽象レベル高 ─── PFC/Executive ノード群 (意思決定・進化戦略制御)
                    │
             ロール割り当て信号
                    ↓
             統合ノード群 (SpatialIntegration, LanMain)
                    │
             特徴圧縮信号
                    ↓
抽象レベル低 ─── 感覚ノード群 (Vision, Audio, Tactile)

3. Extended Cooperative Co-evolution

3.1 Definition and motivation

The existing Phase 6 Coevolution (coevolution.py) deals with competition/cooperation between two to several populations. Enhanced Cooperative Evolution adds:

Asymmetric cooperation: cooperation of specialized populations with different roles
Communication evolution: Evolving the signaling method itself between populations
Cultural evolution: Mechanism by which learned behaviors are propagated within a group.
Ecosystem dynamics: Formation of an ecosystem where multiple niches are interdependent

3.2 Theoretical framework

3.2.1 Formal definition of cooperative coevolution

There are \(K\) populations \(\{P_1, \ldots, P_K\}\), and the fitness of each individual \(g_i^{(k)}\) is determined by interactions with representative individuals of other populations:

\[ F_k(g_i^{(k)}) = f_k\left(g_i^{(k)}, \text{context}_{-k}\right) \]

Here \(\text{context}_{-k} = \{r^{(j)} : j \neq k\}\) (representative set of individuals of other populations).

3.2.2 Role Specialization Theory

Conditions for the emergence of division of roles within a team:

Complementarity condition: When individual \(a\) is good at a task and individual \(b\) is bad at a task, and vice versa, the combined fitness is higher than alone:

\[ F_{\text{team}}(a, b) > \max\left(F(a, a), F(b, b)\right) \]

Role stability: Nash equilibrium condition — no individual can improve by changing roles alone:

\[ F_k(g_k^*, \mathbf{g}_{-k}^*) \geq F_k(g, \mathbf{g}_{-k}^*) \quad \forall g \in P_k \]

3.2.3 Theory of communication evolution

Communication evolution based on Shannon's information theory:

Signal dimension: Communication channel in \(d_{\text{signal}}\) dimension
Sender fitness: Calculated from recipient behavior changes
Receiver fitness: task performance obtained by interpreting the signal

The evolution of signals eventually leads to the emergence of a ``meaningful language'' (similar to the language evolution model of Nowak & Krakauer, 1999).

\[ \text{Vocabulary acquisition rate} \propto \exp\left(-d_H(s, s') / T\right) \]

Here \(d_H\) is the Hamming distance and \(T\) is the "temperature" parameter.

3.2.4 Cultural evolution mechanism (Meme propagation)

Good solutions within a population are propagated as “culture”:

Imitation learning: Stochastically copying the strategy of individuals with higher fitness
Memetic mutation: Maintaining diversity through imitation errors during copying
Cultural Selection: Advantageous memes spread at the group level.

\[ P(\text{meme}_i \to \text{meme}_j) = \frac{F_j}{\sum_k F_k} \cdot (1 - \epsilon) + \epsilon \cdot \mathcal{U} \]

3.2.5 Ecosystem evolutionary dynamics

Multiple population coexistence model by extending the Lotka-Volterra equation:

\[ \frac{dN_i}{dt} = r_i N_i \left(1 - \frac{\sum_j \alpha_{ij} N_j}{K_i}\right) \]

where: - \(N_i\): Size of population \(i\) - \(r_i\): Intrinsic growth rate (= evolution rate) - \(K_i\): Carrying capacity (= resource limit) - \(\alpha_{ij}\): population interaction matrix (strength of competition/cooperation)

3.3 Specific mechanism in EvoSpikeNet

3.3.1 Extensions to existing CoevolutionEngine

Additional features to the current CoevolutionEngine (coevolution.py):

Existing features	Enhanced content
`competitive_evaluation`	Response to asymmetric competition and environmental changes
`cooperative_evaluation`	Adding role specialization mechanism
`speciate_population`	Addition of ecological niche coexistence model
Fixed team size	Dynamic team formation/dissolution

3.3.2 Specialized population architecture

Specialized populations for distributed nodes in EvoSpikeNet:

個体群アーキテクチャ:
┌─────────────────────────────────────────────────┐
│  P_Executive: 高次意思決定戦略の進化              │
│  P_Memory: 記憶アクセスパターンの進化             │
│  P_Perception: 感覚処理効率の進化                │
│  P_Motor: 運動制御戦略の進化                     │
│  P_Communication: ノード間通信プロトコルの進化    │
└─────────────────────────────────────────────────┘
      ↕ 適応度依存相互作用（役割補完性を最大化）

3.3.3 Communication protocol evolution

Treat Zenoh-based message formats as evolution variables:

Message compression level: Accuracy vs. latency trade-off optimization
Transmission frequency: Burst vs. steady dynamic switching
Priority Mapping: Automatic adjustment of message priority according to task importance

3.3.4 Collective intelligence mechanism

Implementation of Stigmergy (indirect cooperation):

Population leaves traces in the environment (genome pool) as “pheromone”
Successful genetic patterns are recorded as "hotspots"
Other individuals preferentially search around hot spots

4. Adaptive Evolution Strategies

4.1 Definition and motivation

Adaptation of evolutionary strategies is a mechanism that adjusts evolutionary parameters (mutation rate, selection method, population size, etc.) in real time in response to changes in the external environment.

Fixed-parameter evolution strategies work in steady-state environments, but performance drops sharply in real-world unsteady environments (changes in sensor noise, changes in task demands, and fluctuations in hardware performance).

4.2 Theoretical framework

4.2.1 Environmental change detection theory

Detection of environmental change points using CUSUM (cumulative sum) test:

\[ S_t = \max\left(0, S_{t-1} + (x_t - \mu_0) - k\right) \]

When the threshold \(h\) is exceeded, a change point is detected and the readaptation of the evolutionary strategy is triggered.

Stability evaluation by curvature of fitness landscape:

\[ \kappa = \frac{\partial^2 F}{\partial \theta^2}\bigg|_{\theta^*} \]

\(\kappa < 0\) (convex): Stable optimal solution → Increased exploitation
\(\kappa > 0\) (concave): unstable local optimum → increased exploration

4.2.2 Dynamic adjustment of exploration-exploitation balance

UCB (Upper Confidence Bound) type balance control:

\[ \text{Exploration\_bias}(t) = C \cdot \sqrt{\frac{\ln t}{n_{\text{exploit}}(t)}} \]

ε-greedy dynamic scheduling:

\[ \epsilon(t) = \epsilon_{\min} + (\epsilon_{\max} - \epsilon_{\min}) \cdot e^{-\lambda t} \]

However, ε is reset when an environmental change is detected: \(\epsilon(t_{\text{change}}) = \epsilon_{\max}\)

4.2.3 CMA-ES (Covariance Matrix Adaptive Evolution Strategy)

Adopt CMA-ES for efficient search in high-dimensional parameter space:

\[ \mathbf{x}^{(g+1)}_k = \mathbf{m}^{(g)} + \sigma^{(g)} \mathcal{N}(\mathbf{0}, \mathbf{C}^{(g)}) \]

\(\mathbf{m}\): mean of distribution (current best estimate)
\(\sigma\): Step size (overall search range)
\(\mathbf{C}\): Covariance matrix (learning search direction)

Updating the covariance matrix (learning the "direction" of evolution):

\[ \mathbf{C}^{(g+1)} = (1-c_1-c_\mu)\mathbf{C}^{(g)} + c_1 \mathbf{p}_c \mathbf{p}_c^T + c_\mu \sum_{i=1}^{\mu} w_i \mathbf{y}_{i:\lambda} \mathbf{y}_{i:\lambda}^T \]

4.2.4 Hierarchical control of global and local adaptation

Combination with Island Model:

Each island (population) has independent evolutionary strategy parameters
Effective strategies spread through regular migration
Global optimization progresses while maintaining inter-island diversity

\[ P_{\text{migrate}}(i \to j) = \sigma\left(\beta \cdot \Delta F_{ij}\right) \]

Here \(\Delta F_{ij} = \bar{F}_j - \bar{F}_i\) (promotion of migration due to fitness difference).

4.2.5 Integration with reinforcement learning (CEM + RL)

Progress configuration system (integration with progress_config.py):

Reinforcement learning optimization of evolutionary strategies using Cross-Entropy Method (CEM):

Initialize the probability distribution \(p(\theta; \mu, \sigma)\) in the evolutionary strategy parameter space \(\Theta\)
Execute episodes with \(\theta\) of \(N\) samples and evaluate fitness
Update \(\mu, \sigma\) from top \(N_e\) elites
Expanding the environmental characteristic vector \(\mathbf{e}\) to the conditional distribution \(p(\theta | \mathbf{e})\)

4.3 Specific mechanism in EvoSpikeNet

4.3.1 Environmental sensor integration

Environmental signals used to adapt evolutionary strategies:

Environmental signals	Data sources	Update frequency	Influencing strategy parameters
Task performance change	`fitness_evaluator.py`	Every generation	Mutation rate, selection pressure
Energy usage rate	`energy_tracker.py`	Real time	Population size, evaluation frequency
Node failure rate	`auto_recovery.py`	Real time	Robustness weight
Input distribution shift	`dataloaders.py`	Batch by batch	Diversity pressure
Network delay	`zenoh_comm.py`	Real time	Distributed evaluation frequency
Memory pressure	`memory_monitor.py`	Periodic	Population size limit

4.3.2 Evolution parameter controller

Adaptive controller extending MutationEngine in evolution_engine.py:

class AdaptiveStrategyController:
    """環境変化に応答して進化戦略を自動調整"""

    def __init__(self, base_config: EvolutionConfig):
        self.current_params = base_config
        self.environment_monitor = EnvironmentMonitor()
        self.history = RollingBuffer(window=50)
        self.change_detector = CUSUMDetector(threshold=5.0)

    def adapt(self, fitness_signal: float, env_state: EnvState):
        """フィットネス信号と環境状態から進化パラメータを更新"""
        self.history.push(fitness_signal)

        if self.change_detector.detect(fitness_signal):
            self._reset_exploration()  # ε reset

        gradient = self._estimate_fitness_gradient()
        self._update_params(gradient, env_state)

    def _update_params(self, gradient: float, env: EnvState):
        if abs(gradient) < self.plateau_threshold:
            # Stagnant state → Enhance exploration
            self.current_params.mutation_rate *= 1.5
            self.current_params.diversity_pressure += 0.1
        else:
            # Under improvement → Strengthen utilization
            self.current_params.elitism_ratio += 0.05
            self.current_params.mutation_rate *= 0.9

4.3.3 Evolution parameters and hardware adaptation

Responding to cluster configuration changes (node addition/deletion):

Increase in number of nodes → proportional increase in population size, increase in number of islands
Node failure → Restoration of affected population from backup
GPU/CPU switching → automatic adjustment of evaluation batch size

5. 4 subsystem integrated architecture

5.1 Integrated model diagram

┌─────────────────────────────────────────────────────────────────┐
│                    11.7 自己進化の高度化                          │
│                                                                   │
│  ┌─────────────┐  フィードバック  ┌─────────────────────────┐   │
│  │ メタ進化    │ ←──────────── │  進化戦略の適応          │   │
│  │ (L4)        │ ──────────→  │  (环境センサー統合)       │   │
│  └──────┬──────┘  戦略制御     └──────────┬──────────────┘   │
│         │ 制約/フィードバック              │ パラメータ調整    │
│         ↓                                 ↓                     │
│  ┌─────────────┐               ┌─────────────────────────┐   │
│  │ 階層的進化  │ ←────────── │  協調進化の拡張          │   │
│  │ (多時間スケ │   個体群間    │  (役割特化+文化的進化)   │   │
│  │  ール)      │   シグナル    │                          │   │
│  └──────┬──────┘               └──────────┬──────────────┘   │
│         │                                 │                     │
│         └──────────┬──────────────────────┘                    │
│                    ↓                                             │
│          ┌─────────────────┐                                    │
│          │ EvoGenome       │                                    │
│          │ (Phase 5/6 基盤) │                                    │
│          └─────────────────┘                                    │
└─────────────────────────────────────────────────────────────────┘

5.2 Signal Flow Priority

Priority	Signal	Source	Destination	Trigger Condition
Best	Safety Violation Alert	`safety_filter.py`	Full Evolution System	Immediate Stop
High	Sudden environmental change signal	`change_detector`	Adaptation of evolutionary strategy	CUSUM threshold exceeded
Medium	Diversity collapse warning	`CoevolutionEngine`	Metaevolution	Number of species < 3
Low	Performance stagnation notification	`EvolutionEngine`	Hierarchical evolution	No improvement for 10 generations

5.3 Organizing computational complexity

Subsystem	Time complexity	Space complexity	Parallelizability
Metaevolution	\(O(G \cdot N \cdot F_{\text{eval}})\)	\(O(N + \\|Θ\\|)\)	Not possible between generations, possible between individuals
Hierarchical evolution	\(O(L \cdot N \cdot F_{\text{eval}})\)	\(O(L \cdot N)\)	Sequential between levels, possible between individuals
Cooperative evolution extension	\(O(K^2 \cdot N^2)\) (at the time of speciation)	\(O(K \cdot N)\)	Possible between populations
Evolutionary strategy adaptation	\(O(W \cdot D^2)\) (CMA-ES)	\(O(D^2)\)	Evaluation can be done in parallel

6. Correspondence with previous research

Concept	Previous research	Implementation on EvoSpikeNet
Metaevolution	CMA-ES (Hansen, 2001), MAML (Finn et al., 2017)	`MetaEvolutionConfig` extension
Hierarchical Evolution	HyperNEAT (Stanley & Miikkulainen), NEAT	`HierarchicalGenome` Structure
Cooperative coevolution	CCEA (Potter & Jong, 1994), NSGA-II	`CoevolutionEngine` extension
Cultural Evolution	MemeticAlgorithms (Moscato, 1989)	`MemeticPropagation` Implementation
Baldwin Effect	Hinton & Nowlan (1987)	`GeneticAssimilation` Module
Evolutionary Strategy Adaptation	IPOP-CMA-ES, Self-Adaptive ES	`AdaptiveStrategyController`

7. EvoSpikeNet integration points with existing architectures

7.1 Extending existing files

Existing files	Expanded content
`evospikenet/genome.py`	`MetaEvolutionConfig`, `HierarchicalChromosome` data class addition
`evospikenet/evolution_engine.py`	`MetaEvolutionEngine`, `AdaptiveStrategyController` classes added
`evospikenet/advanced_mutations.py`	Meta evolution compatible extension of `AdaptiveMutationConfig`
`evospikenet/coevolution.py`	`ExtendedCoevolutionEngine`, `CommunicationEvolutionEngine` added
`evospikenet/genome_pool.py`	Added support for hierarchical population management and ecosystem dynamics

7.2 New file

File	Role
`evospikenet/meta_evolution.py`	Meta evolution engine body
`evospikenet/hierarchical_evolution.py`	Hierarchical evolution controller
`evospikenet/evolution_strategy_adapter.py`	Evolution strategy adaptation controller
`evospikenet/memetic_evolution.py`	Cultural evolution/memetic propagation
`evospikenet/ecosystem_dynamics.py`	Ecosystem dynamics simulation

8. References

Hansen, N. & Ostermeier, A. (2001). Completely Derandomized Self-Adaptation in Evolution Strategies. Evolutionary Computation, 9(2).
Finn, C., Abbeel, P., & Levine, S. (2017). Model-Agnostic Meta-Learning for Fast Adaptation. ICML 2017.
Potter, M.A. & De Jong, K.A. (1994). A Cooperative Coevolutionary Approach to Function Optimization. PPSN III.
Moscato, P. (1989). On Evolution, Search, Optimization, Genetic Algorithms and Martial Arts: Towards Memetic Algorithms. Caltech Memo.
Stanley, K.O. & Miikkulainen, R. (2002). Evolving Neural Networks through Augmenting Topologies. Evolutionary Computation, 10(2).
Hinton, G.E. & Nowlan, S.J. (1987). How Learning Can Guide Evolution. Complex Systems, 1(3).
Wolpert, D.H. & Macready, W.G. (1997). No Free Lunch Theorems for Optimization. IEEE Trans. Evol. Comput., 1(1).
Nowak, M.A. & Krakauer, D.C. (1999). The Evolution of Language. PNAS, 96(14).
Rechenberg, I. (1973). Evolutionsstrategie. Frommann-Holzboog.
Baldwin, J.M. (1896). A New Factor in Evolution. American Naturalist, 30.