Research map / Reinforcement learning / Policy gradient and actor-critic
Implicit Quantile Networks for Distributional Reinforcement 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 |
|---|---|---|---|
| Replace the fixed set of quantiles with an implicit continuous quantile function | DQN, producing IQN · Atari-57, 30 no-op starts | human-normalized mean: 864% -> 1019% | Helped |
| Replace the fixed set of quantiles with an implicit continuous quantile function | DQN, producing IQN · Atari-57, 30 no-op starts | human-normalized median: 193% -> 218% | Helped |
| Use IQN instead of Rainbow | Atari-57 agent · Atari-57, 30 no-op starts | human gap: 0.144 -> 0.141 | Helped |
| Use IQN instead of Rainbow | Atari-57 agent · Atari-57, human-starts | human-normalized median: 153% -> 162% | Helped |
| Increase the number of current-state quantile samples N (ablation) | IQN · Six Atari games, first 10M frames | average human-normalized score | Helped |
| Increase the number of target quantile samples N' (ablation) | IQN · Six Atari games, first 10M frames | average human-normalized score | Helped |
| Increase target quantile samples beyond N' = 8 (ablation) | IQN · Six Atari games, last 10M frames | average human-normalized score | No clear effect |
| Use one sample for each loss distribution (N = N' = 1) (ablation) | IQN · Six Atari games, long-term performance | performance relative to DQN: ≈3× DQN | Helped |
| Use the risk-averse policies reported as beneficial rather than the risk-neutral policy (ablation) | IQN · Asterix and Assault | score under risk-neutral evaluation | Helped |
| Use CVaR(0.1) rather than the risk-neutral policy (ablation) | IQN · QBert and Space Invaders | score under risk-neutral evaluation | Hurt |
| Use CPW rather than the risk-neutral policy (ablation) | IQN · Six Atari games | score under risk-neutral evaluation | No clear effect |
| Use risk-seeking Wang(1.5) rather than the risk-neutral policy (ablation) | IQN · Six Atari games | score under risk-neutral evaluation | Mixed |
Closest papers by meaning
- Distributional Reinforcement Learning With Quantile Regression (AAAI 2018 · similarity 0.93)
- Fully Parameterized Quantile Function for Distributional Reinforcement Learning (NeurIPS 2019 · similarity 0.93)
- Quantile Geometry Regularization for Distributional Reinforcement Learning (NeurIPS 2026 · similarity 0.92)
- Tilted Quantile Gradient Updates for Quantile-Constrained Reinforcement Learning (AAAI 2025 · similarity 0.86)
- A Principled Path to Fitted Distributional Evaluation (NeurIPS 2025 · similarity 0.86)
- A Distributional Analogue to the Successor Representation (ICML 2024 · 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.