1990two/evolutionary_turing

Evolutionary Turing Machine - The Classics Revival

Neural Turing Machines That Evolve Their Own Architectures

Experimental Research Code - Functional but unoptimized, expect rough edges

What Is This?

Evolutionary Turing Machine combines Neural Turing Machines with evolutionary algorithms to create memory-augmented networks that evolve their own architectures. Instead of hand-designing memory systems, populations of NTMs compete and evolve optimal memory configurations.

Core Innovation: NTMs that mutate their memory slots, read/write heads, and controller architectures through evolutionary pressure, discovering novel memory patterns.

Architecture Highlights

• Self-Evolving Memory: Memory slots and dimensions adapt through evolution
• Adaptive Read/Write Heads: Number and behavior of memory heads evolve
• Controller Architecture Search: LSTM controller dimensions discovered evolutionarily
• Task-Driven Fitness: Evolution guided by performance on memory tasks
• Population Diversity: Maintains genetic diversity in memory architectures
• Crossover Operations: Recombines successful memory strategies

Quick Start

from evolutionary_turing import EvolutionaryTuringMachine, EvolutionaryTuringConfig

# Create evolutionary NTM system
config = EvolutionaryTuringConfig(
    population_size=50,
    input_dim=8,
    output_dim=8,
    max_generations=100
)

evolution = EvolutionaryTuringMachine(config)

# Evolve population on memory tasks
history = evolution.run_evolution()

# Get the best evolved model
best_ntm = evolution.get_best_model()

Current Status

• Working: Population evolution, architecture mutations, memory task evaluation, crossover operations
• Rough Edges: No distributed evolution, limited task variety, basic fitness functions
• Still Missing: Advanced mutation operators, multi-objective optimization, neural architecture search integration
• Performance: Functional on toy problems, needs scaling for complex tasks
• Memory Usage: High due to population storage, optimization needed
• Speed: Sequential evaluation, parallelization would help significantly

Mathematical Foundation

The evolutionary process optimizes the NTM architecture space through genetic algorithms:

Fitness(NTM) = Performance(copy_task) × 0.5 + Performance(recall_task) × 0.3 + Efficiency × 0.2

Mutations modify:

• Controller dimensions: d_new = d_old + N(0, σ_controller)
• Memory parameters: M_slots_new ~ U[16, 256], M_dim_new ~ U[8, 64]
• Head configurations: heads_read_new ~ U[1, 4], heads_write_new ~ U[1, 3]

Selection pressure favors architectures that balance task performance with computational efficiency.

Research Applications

• Neural architecture search for memory systems
• Adaptive memory allocation strategies
• Meta-learning through evolutionary computation
• Automated machine learning for sequence tasks
• Evolutionary neural network design

Installation

pip install torch numpy
# Download evolutionary_turing.py from this repo

The Classics Revival Collection

Evolutionary Turing Machine is part of a larger exploration of foundational algorithms enhanced with modern neural techniques:

• Evolutionary Turing Machine ← You are here
• Hebbian Bloom Filter
• Hopfield Decision Graph
• Liquid Bayes Chain
• Liquid State Space Model
• Möbius Markov Chain
• Memory Forest

Citation

@misc{evolutionaryturing2025,
  title={Evolutionary Turing Machine: Self-Evolving Memory Architectures},
  author={Jae Parker 𓅸 1990two},
  year={2025},
  note={Part of The Classics Revival Collection}
}