Research map / Reinforcement learning / Policy gradient and actor-critic
Diverse Exploration for Fast and Safe Policy Improvement
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 |
|---|---|---|---|
| Deploy diverse safe policies instead of a single safe policy | DE versus SPI · Grid World | average aggregate normalized return: 604.970 -> 675.562 | Helped |
| Deploy diverse safe policies instead of a single safe policy | DE versus SPI · Mountain Car | average aggregate normalized return: 362.038 -> 381.333 | Helped |
| Deploy diverse safe policies instead of a single safe policy | DE versus SPI · Acrobot | average aggregate normalized return: 417.145 -> 430.146 | Helped |
| Deploy diverse safe policies instead of a single safe policy | DE versus SPI · Grid World | iterations to confirm at least one optimal policy: 22 -> 16 | Helped |
| Deploy diverse safe policies instead of a single safe policy | DE versus SPI · Grid World | successful runs confirming an optimal policy: DE ... all 50 runs; SPI ... unsuccessful in 5 runs | Helped |
| Use multiple importance sampling with equal trajectory allocation across safe policies | DE · Theoretical estimator comparison | uniformity of estimator variances | Helped |
| Use equal allocation across behavior policies rather than selecting one policy for all trajectories | DE · Theoretical estimator comparison | average variance across estimators | No clear effect |
| Use a large mixing parameter for candidate policies | Mountain Car and Acrobot policy generation · Mountain Car and Acrobot | iterations required for DE to outperform SPI | Mixed |
| Bootstrap separately within trajectories collected per iteration | DE candidate policy generation · Policy generation | representation of policies in bootstrapped training data | Helped |
Closest papers by meaning
- DGPO: Discovering Multiple Strategies with Diversity-Guided Policy Optimization (AAAI 2024 · similarity 0.89)
- Robust Policy Expansion for Offline-to-Online RL under Diverse Data Corruption (NeurIPS 2025 · similarity 0.86)
- Balance Reward and Safety Optimization for Safe Reinforcement Learning: A Perspective of Gradient Manipulation (AAAI 2024 · similarity 0.86)
- Learning Diverse Risk Preferences in Population-Based Self-Play (AAAI 2024 · similarity 0.86)
- Scalable Online Exploration via Coverability (ICML 2024 · similarity 0.86)
- Controlling Behavioral Diversity in Multi-Agent Reinforcement Learning (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.