Reference library
Deep Neuroevolution: Genetic Algorithms Are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning
Paper summary
A simple Deep GA evolves fixed-topology neural parameters using truncation selection, additive Gaussian mutations and elitism, omitting crossover for simplicity. A network with more than four million parameters achieves competitive results on selected Atari benchmarks. The paper discusses unequal frame budgets and evaluation-start conventions among some external comparisons. This establishes a useful mutation-only baseline, but contains no ablation showing crossover would worsen that GA.
Relevance to the project
Assistant interpretation: mutation-only evolution is a substantive baseline when assessing whether recombination of inherited neural rules earns its complexity and evaluation cost. Fixed action-policy weights are distinct from lifetime-plastic F/H weights and inherited type controllers.
Older benchmark-specific comparisons, five Atari runs and heterogeneous external settings; the current review does not audit all experiments or supplement. Success without recombination does not establish its inferiority.
Source record
- Reference ID
- deep-neuroevolution-2018
- Archived version
- arXiv:1712.06567v3, 20 April 2018
- Summary provenance
- reference-library/papers/deep-neuroevolution-2018/metadata.json · summary
- Review status
- Targeted primary full-text review; not exhaustive reading or reproduction
- Reading scope
- Targeted PDF pp.2–5: sections3.1,4 and4.1 on mutation-only GA, Atari setup/results and external-comparison caveats. Supplement/locomotion/novelty experiments not fully audited; no independent reproduction.