Equivariant RL
summary
Argues that encoding symmetry cuts sample complexity: policies built with group-equivariant convolutions and equivariant graph neural networks learn from far fewer environment interactions and show returns rising much faster on benchmarks like cart‑pole and symmetric multi‑agent tasks. Implements 4‑way (90°) rotation capsules and SE3/spherical‑harmonic variants for continuous rotations, links MDP homomorphisms to orbits, and reports that hard‑coded equivariance outperforms plain data augmentation while allowing approximate equivariance via a penalty or sampled orbit enforcement.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:00What is RL
- 19:38Advantages of Equivariant
- 26:42Question
- 39:50Results
- 40:29Potential gains
- 45:15Experiments
- 50:58Graph Neural Networks
- 56:13Conclusion
related talks
01h 29mBoot camp on invariances in graph learning
01h 04mRandomized Embeddings and Neural Networks
54mLearning Uninformative Representations
01h 11m(Hidden) Symmetries of Computational Problems (continued)
32mEnd-to-End Optimizing Multi-Turn RL and High-Performance Inference in Agents with...
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
26mApplying contrastive learning to fight anisotropy in language models
34mOffline Reinforcement Learning and Model-Based Optimization