Research map / Reinforcement learning / Visual and representation RL
Deep Reinforcement Learning That Matters
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 DDPG rather than TRPO | policy-gradient baselines · HalfCheetah-v1 | return: 5037.26 vs 1254.55 | Helped |
| Use TRPO rather than DDPG | policy-gradient baselines · Hopper-v1 | return: 2965.33 vs 1632.13 | Helped |
| Use TRPO rather than DDPG | policy-gradient baselines · Walker2d-v1 | return: 3072.97 vs 1582.04 | Helped |
| Use TRPO rather than DDPG | policy-gradient baselines · Swimmer-v1 | return: 214.69 vs 31.92 | Helped |
| Change the policy or value network architecture or activation (ablation) | policy-gradient methods · Hopper-v1 and HalfCheetah-v1 | return | Mixed |
| Use ReLU rather than tanh activations (ablation) | policy-gradient methods · Hopper-v1 and HalfCheetah-v1 | return | Mixed |
| Change DDPG reward scaling (ablation) | DDPG · MuJoCo continuous-control environments | return | Mixed |
| Use reward scaling below 0.01 (ablation) | DDPG · HalfCheetah-v1 | return: 0.01 | Hurt |
| Use a TRPO batch size of 1024 rather than larger batch sizes (ablation) | TRPO · Hopper-v1 and HalfCheetah-v1 | return: 1024 | Helped |
| Use different random seeds with identical hyperparameters (ablation) | TRPO and DDPG · HalfCheetah-v1 and other tested environments | return: t = -9.0916, p = 0.0016 | Mixed |
| Change the implementation codebase while holding hyperparameters constant (ablation) | TRPO and DDPG · Hopper-v1 and HalfCheetah-v1 | return: 2021.34 vs 2965.3; 3576.08 vs 1045.6 | Mixed |
Closest papers by meaning
- Deep Reinforcement Learning at the Edge of the Statistical Precipice (NeurIPS 2021 · similarity 0.90)
- Beyond Expected Return: Accounting for Policy Reproducibility When Evaluating Reinforcement Learning Algorithms (AAAI 2024 · similarity 0.89)
- Position: Benchmarking is Limited in Reinforcement Learning Research (ICML 2024 · similarity 0.86)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.86)
- XLand-100B: A Large-Scale Multi-Task Dataset for In-Context Reinforcement Learning (ICLR 2025 · similarity 0.86)
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · 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.