Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Natural Evolution Strategies

Authors: Daan Wierstra, Tom Schaul, Tobias Glasmachers, Yi Sun, Jan Peters, Jürgen Schmidhuber

JMLR 2014 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically validate the new algorithms, to determine how NES algorithms perform compared to state-of-the-art evolution strategies, identifying specific strengths and limitations of the different variants. We conduct a broad series of experiments on standard benchmarks, as well as more specific experiments testing special capabilities.
Researcher Affiliation Collaboration Daan Wierstra EMAIL Tom Schaul EMAIL Deep Mind Technologies Ltd. Fountain House, 130 Fenchurch Street London, United Kingdom Tobias Glasmachers EMAIL Institute for Neural Computation Universit atsstrasse 150 Ruhr-University Bochum, Germany Yi Sun EMAIL Google Inc. 1600 Amphitheatre Pkwy Mountain View, United States Jan Peters EMAIL Intelligent Autonomous Systems Institute Hochschulstrasse 10 Technische Universit at Darmstadt, Germany J urgen Schmidhuber EMAIL Istituto Dalle Molle di Studi sull Intelligenza Artificiale (IDSIA) University of Lugano (USI)/SUPSI Galleria 2 Manno-Lugano, Switzerland
Pseudocode Yes Algorithm 1: Canonical Search Gradient algorithm Algorithm 2: Search Gradient algorithm: Multinormal distribution Algorithm 3: Canonical Natural Evolution Strategies Algorithm 4: Adaptation sampling Algorithm 5: Exponential Natural Evolution Strategies (x NES) (multinormal case) Algorithm 6: Separable NES (SNES)
Open Source Code Yes A Python implementation of all these is available within the open-source machine learning library Py Brain (Schaul et al., 2010), and implementations in different languages can be found at http://www.idsia.ch/~tom/nes.html.
Open Datasets Yes We evaluate our algorithm on all the benchmark functions of the Black-Box Optimization Benchmarking collection (BBOB) from the GECCO Workshop for Real-Parameter Optimization. The collection consists of 24 noise-free functions (12 unimodal, 12 multimodal; Hansen et al., 2010a) and 30 noisy functions (Hansen et al., 2010b).
Dataset Splits No The paper uses benchmark functions like the BBOB collection and a neuroevolution task. It describes evaluation criteria such as a "budget of function evaluations (10^5d)" and reaching a "target value fopt + 10k" for these benchmarks, but it does not specify explicit training/test/validation dataset splits in the conventional sense for a static dataset.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running experiments.
Software Dependencies No A Python implementation of all these is available within the open-source machine learning library Py Brain (Schaul et al., 2010). No specific version numbers for Python or Py Brain are provided.
Experiment Setup Yes Table 1: Default parameter values for x NES, x NES-as and SNES (including the utility function) as a function of problem dimension d. This table provides specific hyperparameters such as λ, ηµ, ησ, ηB.