Source-linked AI summary
Evolving neural networks with genetic algorithms to study the String Landscape
Fabian Ruehle
TL;DR
The paper asks how machine learning can analyze extensive string-landscape data despite unclear neural-network architectures and costly computations. It evolves neural networks with genetic algorithms for classification, known-but-tedious computations, and approximation of inefficient computations. In the line-bundle-cohomology example, the evolved networks achieve task-dependent accuracy and reduce computation from several hours to 7 seconds.
Problem
String theory contains extensive landscape data, but machine-learning analysis is undeveloped, model selection lacks a clear mechanism, and suitable neural-network architectures are unclear for different tasks.
Method
The paper evolves neural networks through genetic-algorithm reproduction and mutation, applying them to classification, known computations, and regression-based approximation.
Results
The evolved networks predict h0(L) and h3(L) correctly in 82 percent of cases, h1(L) and h2(L) in 73 percent, and compute cohomology dimensions in 7 seconds instead of several hours exactly.
Takeaways & Limitations
Evolving neural networks can provide useful string-phenomenology models for classification, computation, and fast approximation of costly calculations.
Takeaways & Limitations
Performance depends on task formulation and training choices: using prediction for a classification task produced 90 percent accuracy, while inappropriate training can undertrain or overtrain the network.
Abstract
from arXiv · showhide
We study possible applications of artificial neural networks to examine the string landscape. Since the field of application is rather versatile, we propose to dynamically evolve these networks via genetic algorithms. This means that we start from basic building blocks and combine them such that the neural network performs best for the application we are interested in. We study three areas in which neural networks can be applied: to classify models according to a fixed set of (physically) appealing features, to find a concrete realization for a computation for which the precise algorithm is known in principle but very tedious to actually implement, and to predict or approximate the outcome of some involved mathematical computation which performs too inefficient to apply it, e.g. in model scans within the string landscape. We present simple examples that arise in string phenomenology for all three types of problems and discuss how they can be addressed by evolving neural networks from genetic algorithms.
1 Introduction
The paper targets the lack of machine-learning methods for analyzing extensive string-theory data and the absence of a clear mechanism for selecting among many vacua. It proposes neural networks, evolved with genetic algorithms, for classification, computation, and prediction tasks in the string landscape.
- Motivation: String theory has accumulated extensive data, but machine-learning techniques had not yet been applied to analyze it or select among its many vacua.The motivating challenge includes large databases of Calabi–Yau manifolds, string models, and flux vacua.
- Applications: Neural networks can classify models, represent known input-output computations, or approximate costly computations such as bundle cohomology calculations.These correspond to classification, computation, and regression-based prediction applications.
- Applications: The universal approximation theorem supports using sufficiently complex neural networks to approximate functions without manually specifying an interpolation function or selected input features.This motivates trainable neural networks for involved or inefficient computations.
- Method: Because the best architecture is unclear and depends on the task, the paper evolves neural networks with genetic algorithms to optimize their performance.Architecture choices include layer number, layer sizes, and layer types; classification and regression may favor different layers.
- Scope: The paper presents pedagogical neural-network examples, introduces genetic algorithms for breeding effective networks, and applies the approach to predicting line bundle cohomology.The paper also notes that neural-network and genetic-algorithm computations can be parallelized.
2 Artificial neural networks
Artificial neural networks map inputs to outputs through modular layers, making them suitable for classification, regression, and involved string-theory computations. The section introduces layer-based representations and examples involving Euler numbers, line-bundle cohomology, bundle stability, and model classification.
- ANN structure: An ANN maps input data to output data through connected layers, whose perceptrons apply linear transformations and activation functions.Networks may also branch or include feedback, producing recurrent architectures.
- Layer representation: Separating connection layers from function layers gives a modular representation that can encode topology, branches, and reusable network building blocks.Linear layers describe connections, while function layers apply specified functions to their inputs.
- Examples: For Euler-number prediction, a single linear layer learns χ = 2(h1,1 − h2,1) from two Hodge numbers.After 30 training rounds, taking about .3 seconds on a regular laptop, the network learned the relation with high accuracy.
- Examples: Line-bundle cohomology can be implemented with layers performing additions, multiplications, and zeroing operations, but the required computation becomes tedious as its steps accumulate.This motivates evolving networks for more complicated regression tasks.
- Examples: A bundle-stability classifier trained on 3 000 of 10 000 line bundles evaluates 7 000 held-out models, illustrating a nontrivial ANN topology.The example uses integer inputs and binary stability labels.
- Examples: Using a predictor for binary stability produced 90 percent accuracy and strange boundary behavior, whereas normalization enabled correct classification in all cases after a few seconds.The mismatch arose because the predictor approximated arbitrary labels 0 and 1 rather than directly encoding class likelihood.
3 Genetic algorithms
The paper evolves neural networks with genetic algorithms by assembling functional components, training candidate populations, and iteratively selecting, reproducing, and mutating the fittest individuals. This approach targets classification, known-but-tedious computations, and costly mathematical predictions in string phenomenology.
- Genetic algorithms evolve neural networks from small functional components into task-specific architectures.The gene pool can contain individual layers or combinations of layers that perform operations such as component-wise vector multiplication.
- Each generation combines gene-pool elements into ANN individuals, trains them, evaluates performance on unknown data, and carries the fittest forward.The process repeats through reproduction and mutation, with fitness determining which individuals pass their genes to the next generation.
- Mutation can add, remove, or replace genes by inserting ANN components at different positions in the network.The procedure cuts the network between genes and splices in a new gene, including at the beginning or end.
- Success depends on evolutionary and training choices, including mutation rate and training duration, because poor settings can limit exploration or distort fitness.Low mutation rates may prevent discovery of substantially different individuals, while undertraining or overtraining can favor unsuitable architectures or training-set behavior.
- For predictor ANNs, the method can approximate expensive computations after training on a subset of examples, including costly string-model calculations.Examples include computations involving Groebner bases, Diophantine equations, and resultants, as well as line bundle cohomologies.
- The framework supports classification when the relevant computation or efficient ANN mapping is unclear, allowing architectures to evolve toward better classification performance.The same evolutionary reasoning is applied when the computational steps are unknown or difficult to map efficiently onto an ANN.
4 Example: Evolving an ANN to compute bundle cohomology
The paper evolves four neural-network species to predict the four line bundle cohomology dimensions independently, using genetic selection over network building blocks. The evolved networks achieve dimension-dependent accuracy, accelerate cohomology evaluation substantially, and transfer to another Calabi–Yau after retraining.
- Setup: The training data contain 2,401 line bundles, divided into 600 training examples and 1,800 validation examples for a complete intersection Calabi–Yau threefold.
- Setup: The study evolves four neural-network species independently, one for each cohomology dimension h_i(L), using genetic algorithms to select network architectures.Each generation retains the two fittest individuals, which produce seven children; the initial generation uses up to four randomly selected gene-pool elements.
- Results: 82 percent of h0(L) and h3(L) predictions are correct, compared with 73 percent for h1(L) and h2(L).The difference is attributed most likely to the former groups being zero more often and the latter groups receiving more non-trivial contributions.
- Results: 7 seconds are required for the trained ANNs to compute bundle cohomology dimensions, compared with several hours for exact Koszul- and Leray-based computations.
- Results: Maximal and average fitness stop increasing significantly after roughly six or seven generations, while training beyond 60 seconds does not improve performance.The authors suggest that the supplied gene pool may limit the fittest architecture attainable in this setup.
- Applicability to other environments: Networks evolved on Calabi–Yau 6784 perform equally well on Calabi–Yau 7862 after retraining with data from the new environment.The two manifolds differ in codimension and projective ambient space but share the same four-dimensional input format.