Reference library
Policy Optimization by Genetic Distillation
Paper summary
GPO combines selected parents in behavior space: learn a state-dependent parent selector, then train a child by behavioral cloning with DAgger. Mutation uses policy gradients, optionally sharing samples. Nine MuJoCo tasks and two-task ablations support the combined method; data sharing supplies the largest isolated gain, with crossover contributing further. Omitting crossover in the ablation copies the stronger parent. The timestep accounting includes imitation rollouts.
Relevance to the project
Assistant interpretation: provides an alternative when inherited controller components cannot be meaningfully aligned. Applying it would require an explicit behavior objective/dataset and a training budget, and would not be classical genetic splicing or an adopted ecology learning mechanism.
Hybrid gradient learning/imitation, fixed policy architecture and externally rewarded control tasks. Not a demonstration of gradient-free weight crossover, spontaneous ecological selection, or superiority at matched total compute.
Source record
- Reference ID
- policy-optimization-genetic-distillation-2018
- Archived version
- arXiv:1711.01012v2, 12 March 2018; ICLR2018 conference paper
- Summary provenance
- reference-library/papers/policy-optimization-genetic-distillation-2018/metadata.json · summary
- Review status
- Targeted primary full-text review; not exhaustive reading or reproduction
- Reading scope
- Targeted PDF pp.3–10, sections3.1–3.3 and4.1–4.4 on genetic operators, evaluation and component ablations; Figure 5 visually checked on p.10. Supplementary A2C results, all hyperparameters and raw trajectories not exhaustively audited; no reproduction.