Research map / Reinforcement learning / Policy gradient and actor-critic
Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a quasimetric QRL objective instead of Q-Learning | MountainCar discretized · reach 9 states | normalized return: 85.53 ± 3.69 vs 85.55 ± 3.57 | No clear effect |
| Use asymmetric quasimetric distances rather than symmetric ℓ2 distances (ablation) | QRL · MountainCar discretized, reach 9 states | normalized return: 73.27 ± 0.84 -> 85.55 ± 3.57 | Helped |
| Use asymmetric quasimetric distances rather than symmetric ℓ2 distances (ablation) | QRL · MountainCar discretized, reach top of hill | normalized return: 95.42 ± 0.16 -> 97.69 ± 0.26 | Helped |
| Replace QRL's optimal-value objective with a contrastive objective (ablation) | Quasimetric Q-function on MountainCar · reach 9 states | normalized return: 72.28 ± 4.63 -> 85.55 ± 3.57 | Helped |
| Replace QRL's optimal-value objective with a contrastive objective (ablation) | Quasimetric Q-function on MountainCar · reach top of hill | normalized return: 83.90 ± 8.73 -> 97.69 ± 0.26 | Helped |
| Use QRL instead of Contrastive RL | D4RL maze large · multi-goal | normalized episode return: 172.64 ± 5.13 -> 187.71 ± 7.62 | Helped |
| Use QRL instead of MSG | D4RL maze umaze · single-goal | normalized episode return: 101.1 ± 26.3 -> 71.72 ± 26.21 | Hurt |
| Use QRL instead of Contrastive RL | D4RL maze medium · multi-goal | normalized episode return: 137.01 ± 6.26 -> 150.51 ± 3.77 | Helped |
| Use QRL instead of Contrastive RL | D4RL maze umaze · multi-goal | normalized episode return: 142.43 ± 11.99 -> 150.60 ± 5.32 | Helped |
| Use QRL instead of Contrastive RL | Online FetchPush · state-based | sample efficiency: 4.9× | Helped |
| Use QRL instead of Contrastive RL | Online FetchSlide · state-based | sample efficiency: 2.7× | Helped |
Closest papers by meaning
- Offline Goal-conditioned Reinforcement Learning with Quasimetric Representations (NeurIPS 2025 · similarity 0.91)
- QPRL : Learning Optimal Policies with Quasi-Potential Functions for Asymmetric Traversal (ICML 2025 · similarity 0.90)
- Minimax Optimal Reinforcement Learning with Quasi-Optimism (ICLR 2025 · similarity 0.88)
- A Nearly Optimal and Low-Switching Algorithm for Reinforcement Learning with General Function Approximation (NeurIPS 2024 · similarity 0.87)
- Model-based RL as a Minimalist Approach to Horizon-Free and Second-Order Bounds (ICLR 2025 · similarity 0.87)
- DARLING: Detection Augmented Reinforcement Learning with Non-Stationary Guarantees (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.