Neuroevolution — LLMpedia

Neuroevolution
Name	Neuroevolution
Caption	Evolutionary synthesis of artificial neural architectures
Specialty	Artificial intelligence, evolutionary computation, machine learning

Contents

Neuroevolution is the application of evolutionary computation techniques to optimize artificial neural network architectures, connection weights, and learning rules. It combines ideas from John von Neumann, Alan Turing, Norbert Wiener, IBM, and Bell Labs research traditions with techniques from Darwinism-inspired algorithms such as those developed by John Holland, David E. Goldberg, and Ingo Rechenberg. Neuroevolution has influenced projects at MIT, Stanford University, Carnegie Mellon University, Google DeepMind, and OpenAI.

Overview

Neuroevolution uses population-based search processes similar to systems by John Holland, Geoffrey Hinton, Yann LeCun, Jürgen Schmidhuber, and Marvin Minsky to evolve network topologies and synaptic weights. Early approaches drew on principles from Evolutionary Strategies, Genetic Algorithms established at University of Michigan and University of Illinois Urbana-Champaign, while modern methods intersect with techniques advanced at University of California, Berkeley, University of Toronto, DeepMind Technologies, and Microsoft Research. Implementations often integrate software ecosystems and toolkits originating from TensorFlow, PyTorch, Theano, Scikit-learn, and standards promoted at conferences such as NeurIPS, ICML, ICLR, AAAI, and GECCO.

Foundational work linked to Alan Turing's ideas and to optimization research at RAND Corporation set context for neuroevolution. The formalization of genetic algorithms by John Holland and practical advances by David E. Goldberg and Ingo Rechenberg spurred adaptation to neural models at labs including Bell Labs and universities like University of Illinois and University of Michigan. Milestones include topological evolution methods emerging from research groups led by figures associated with University of Texas at Austin and University of Sheffield, and the introduction of speciation and complexification ideas popularized in projects connected to Stanford University and SRI International. Later integration with deep learning paradigms occurred in collaborations among Google, DeepMind, OpenAI, Facebook AI Research, and academic teams at Massachusetts Institute of Technology and Princeton University.

Core algorithms extend paradigms from John Holland and David E. Goldberg with adaptations by researchers linked to Stanford University, University of Texas, and SRI International. Representative families include weight-evolution schemes influenced by Evolutionary Strategies and Covariance Matrix Adaptation developed by groups at ETH Zurich and INRIA, topology-evolving frameworks such as neuroevolution of augmenting topologies pioneered in work associated with Stanford University and researchers connected to George Mason University, and indirect encodings inspired by developmental biology research at Harvard University and California Institute of Technology. Techniques incorporate crossover and mutation operators refined in studies at University College London and Imperial College London, multiobjective formulations advanced at University of Amsterdam and Tokyo Institute of Technology, and surrogate-assisted optimization methods from Argonne National Laboratory and Lawrence Berkeley National Laboratory. Hybridizations combine gradient-based optimizers used at Google, Microsoft Research, and University of Toronto with evolutionary search schemas trialed at IBM Research and Intel Labs.

Neuroevolution has been applied to control problems in domains researched at NASA, European Space Agency, and JAXA, robotics platforms developed at Boston Dynamics and Honda Research Institute, game-playing systems evaluated in tournaments organized by ACM and IEEE, and automated design tasks used in industrial settings at Siemens and General Electric. It appears in research on autonomous vehicles at Tesla, Waymo, and Uber ATG, adaptive signal processing efforts at Bell Labs and Nokia Bell Labs, and in computational creativity projects affiliated with MIT Media Lab and Queen Mary University of London. Neuroevolution supports applications ranging from financial modeling in firms like Goldman Sachs and JPMorgan Chase to bioinformatics collaborations at National Institutes of Health and European Molecular Biology Laboratory.

Benchmarking draws on standardized tasks and suites cultivated by communities associated with OpenAI Gym, Atari Competition organizers, and datasets from ImageNet, CIFAR-10, MNIST, and UCI Machine Learning Repository. Comparative studies reference performance baselines set by teams at DeepMind, Google Brain, Facebook AI Research, and OpenAI and leverage frameworks and leaderboards maintained by Papers with Code, Kaggle, and GitHub. Empirical assessments often use metrics and experimental protocols popularized at conferences such as NeurIPS, ICML, and ICLR and datasets curated by institutions like Stanford University and University of California, Irvine.

Practical challenges intersect with compute resource debates involving NVIDIA, AMD, Intel Corporation, and supercomputing centers at Oak Ridge National Laboratory and Lawrence Livermore National Laboratory. Reproducibility concerns echo discussions at Nature, Science (journal), and community standards advocated by organizers of NeurIPS and ICML. Ethical considerations overlap with policy and governance dialogues at European Commission, United Nations, World Economic Forum, and Federal Trade Commission and with safety research undertaken by OpenAI, DeepMind Safety Research, and academic groups at Oxford University and Cambridge University. Issues include environmental impacts debated in forums like COP26, fairness and bias matters engaged by ACM Conference on Fairness, Accountability, and Transparency, and dual-use risk assessments in reports from RAND Corporation and Brookings Institution.