Research map / Reinforcement learning / Visual and representation RL
Learning to Identify Critical States for Reinforcement Learning from Videos
Haozhe Liu, Mingchen Zhuge, Bing Li, Yuhui Wang, Francesco Faccio, Bernard Ghanem, Jürgen Schmidhuber
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 |
|---|---|---|---|
| add compactness loss and reverse loss to importance preservation loss (ablation) | Deep State Identifier detector · Grid World-S | F-1 score (%): 68.98 -> 80.28 | Improved |
| add importance preservation loss to compactness + reverse loss (ablation) | Deep State Identifier detector · Grid World-S | F-1 score (%): 74.42 -> 80.28 | Improved |
| add compactness loss to imp + rev loss (ablation) | Deep State Identifier detector · Grid World-S | F-1 score (%): 76.09 -> 80.28 | Improved |
| add reverse loss to imp + compactness loss (ablation) | Deep State Identifier detector · Grid World-S | F-1 score (%): unstable -> 80.28 | Improved |
| use state-only input (no actions or policy) instead of EDGE with actions and policy parameters | Atari-Pong-S attack · Atari-Pong-S | win rate change: -65.47 -> -77.67 | Improved |
| train detector on multiple policies instead of a single policy (ablation) | Deep State Identifier · Atari-Pong cross-policy (Arch.) | win rate change: -39.55 -> -76.50 | Improved |
| single-policy training evaluated on unseen policies versus random-state baseline (ablation) | Deep State Identifier · Atari-Pong cross-policy (seeds) | win rate change: -63.32 -> -30.67 | Hurt |
| set DQN lookahead steps from detected critical states | Multi-step DQN · Atari-Seaquest | return: 1987.00 -> 4147.25 | Improved |
| critical-state-based n-step instead of grid-searched n-step | Multi-step DQN · Atari-Seaquest | return: 3936.50 -> 4147.25 | Improved |
| use key-conditioned states (opening door with key) | detector confidence · GridWorld-S | confidence score: 67.13 -> 80.65 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Edge — EDGE: Explaining Deep Reinforcement Learning Policies (NeurIPS 2021) · better on 1 of 1 shared result
- Rudder — RUDDER: Return Decomposition for Delayed Rewards (NeurIPS 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- StateMask: Explaining Deep Reinforcement Learning through State Mask (NeurIPS 2023 · similarity 0.87)
- Discovering symbolic policies with deep reinforcement learning (ICML 2021 · similarity 0.85)
- Local Explanations for Reinforcement Learning (AAAI 2023 · similarity 0.85)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.85)
- Bridging State and History Representations: Understanding Self-Predictive RL (ICLR 2024 · similarity 0.83)
- DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization (ICLR 2024 · similarity 0.83)
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.
