Research map / Reinforcement learning / Policy gradient and actor-critic
Deep Exploration via Bootstrapped DQN
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 |
|---|---|---|---|
| Bootstrapped DQN vs DQN on AUC-100 average | Bootstrapped DQN · Atari 14 games | AUC-100: 0.29 -> 0.62 | Helped |
| Remove gradient rescaling by 1/K (Boot-DQN+ vs Boot-DQN) (ablation) | Bootstrapped DQN · Atari 14 games | AUC-100: 0.41 -> 0.62 | Helped |
| Normalize shared-network gradients by 1/K (ablation) | Bootstrapped DQN · Atari | final score / early learning | Mixed |
| Increase number of bootstrap heads K (ablation) | Bootstrapped DQN · Breakout | cumulative reward | Helped |
| Vary data-sharing probability p (0.5 vs 1) (ablation) | Bootstrapped DQN · Atari | cumulative reward | No clear effect |
| Bootstrapped DQN replaces epsilon-greedy exploration | DQN · Atari 49 games | frames to human performance: 30% faster | Helped |
| Bootstrapped DQN vs DDQN on peak score | Bootstrapped DQN · Atari 49 games | max evaluation score | No clear effect |
| Bootstrapped DQN vs DQN on Bowling AUC-100 | Bootstrapped DQN · Atari Bowling | AUC-100: 0.24 -> 0.56 | Helped |
| Bootstrapped DQN vs DQN on Breakout peak score (DDQN) | Bootstrapped DQN · Atari Breakout | max evaluation score: 437.0 -> 855.0 | Helped |
| Bootstrapped DQN vs DDQN on Seaquest peak score | Bootstrapped DQN · Atari Seaquest | max evaluation score: 19183.9 -> 9083.1 | Hurt |
| Bootstrapped DQN on Montezuma's Revenge | Bootstrapped DQN · Atari Montezuma Revenge | max evaluation score: 0.0 -> 100.0 | Helped |
| Bootstrapped DQN vs DQN on Pong AUC-100 | Bootstrapped DQN · Atari Pong | AUC-100: -0.13 -> 0.02 | Helped |
Closest papers by meaning
- Munchausen Reinforcement Learning (NeurIPS 2020 · similarity 0.90)
- HyperDQN: A Randomized Exploration Method for Deep Reinforcement Learning (ICLR 2022 · similarity 0.88)
- Exploration by random network distillation (ICLR 2019 · similarity 0.87)
- Random Latent Exploration for Deep Reinforcement Learning (ICML 2024 · similarity 0.87)
- Exploration and Anti-Exploration with Distributional Random Network Distillation (ICML 2024 · similarity 0.85)
- Stop Regressing: Training Value Functions via Classification for Scalable Deep RL (ICML 2024 · similarity 0.85)
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.