Reference library
Safe Mutations for Deep and Recurrent Neural Networks through Output Gradients
Paper summary
Safe mutation scales perturbations using change in outputs on sampled parental experience. SM-R uses a line search along a random direction; SM-G scales weights by output sensitivity, with recurrent sensitivities propagated through time. These are output gradients rather than objective/reward gradients. The GA has no crossover, tournament size five, and separately tuned mutation variants. Results vary by task and variant: SM-R performs poorly on the difficult maze and SM-G-SUM provides no advantage over control there. Avoiding additional environment interactions still entails neural forward/backward computation.
Relevance to the project
Assistant proposal: behavioral perturbation control could supplement genotype checks for inherited controllers. A trajectory probe would need to include plasticity and recurrent state, not only instantaneous outputs; this is not an adopted operator.
Safety is relative to the sampled experience and local approximation. Rare behavior, long horizons and changing environments can escape the measure. Derivatives require a suitable execution model; no general open-ended-evolution guarantee.
Source record
- Reference ID
- safe-mutations-output-gradients-2018
- Archived version
- arXiv:1712.06563v3, 1 May 2018
- Summary provenance
- reference-library/papers/safe-mutations-output-gradients-2018/metadata.json · summary
- Review status
- Targeted primary full-text review; not exhaustive reading or reproduction
- Reading scope
- Targeted PDF pp.3–7, with p.5 separately reread, covering behavioral divergence, SM-R/SM-G methods, recurrent sensitivity and benchmark results/discussion; opening identity checked. Appendix, complete tuning audit, code and raw data not reviewed. Figures not visually inspected; no reproduction.