Research map / Reinforcement learning / Policy gradient and actor-critic
Efficient Exploration via Epistemic-Risk-Seeking Policy Optimization
Brendan O'Donoghue
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 epistemic-risk-seeking policy optimization instead of vanilla actor-critic | DeepSea · DeepSea | maximum depth solved: 14 -> 100 | Improved |
| Use epistemic-risk-seeking policy optimization instead of Bootstrapped DQN | DeepSea · DeepSea | maximum depth solved: 50 -> 100 | Improved |
| Add simple optimism bonuses to actor-critic | actor-critic · DeepSea | maximum depth solved: approximately depth 50 | Improved |
| Add replay with prioritized sampling and off-policy corrections | ERSAC · DeepSea | data efficiency: about 4× | Improved |
| Add Gaussian noise to reward targets in replay (ablation) | ERSAC + replay · DeepSea | performance | Improved |
| Add replay data (ablation) | ERSAC · DeepSea, depth 100 | seeds reaching goal: 99 times out of 100 -> 93 times | Hurt |
| Learn the risk-seeking parameter τ rather than fix it (ablation) | ERSAC · DeepSea | robustness across initial τ choices | Improved |
| Use TD-λ values of at least 0.6 (ablation) | ERSAC · DeepSea | instances solved to depth 100: λ ≥ 0.6 | Improved |
| Use longer TD-λ rollouts rather than very small rollouts (ablation) | ERSAC · DeepSea | exploration performance: 1 -> 25 | Improved |
| Add the risk-seeking objective to replay actor-critic | Atari 57 · Atari 57 | frames to reach baseline peak median performance: 1.8× fewer | Improved |
| Add the risk-seeking objective to replay actor-critic | hard-exploration Atari games · 7 hard-exploration Atari games | game performance: four of the 7 games | Mixed |
Closest papers by meaning
- Exploration from a Primal-Dual Lens: Value-Incentivized Actor-Critic Methods for Sample-Efficient Online RL (NeurIPS 2025 · similarity 0.91)
- Exploration via Epistemic Value Estimation (AAAI 2023 · similarity 0.91)
- Beyond Optimism: Exploration With Partially Observable Rewards (NeurIPS 2024 · similarity 0.90)
- NeoRL: Efficient Exploration for Nonepisodic RL (NeurIPS 2024 · similarity 0.89)
- MaxInfoRL: Boosting exploration in reinforcement learning through information gain maximization (ICLR 2025 · similarity 0.88)
- Random Latent Exploration for Deep Reinforcement Learning (ICML 2024 · similarity 0.88)
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.
