Research map / Reinforcement learning / Visual and representation RL
Information Optimization and Transferable State Abstractions in Deep Reinforcement Learning
Diego Gomez, Nicanor Quijano, Luis Felipe Giraldo
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 information concepts with TRMC instead of SAC | SAC · AntSquareWall | episodes to 50% of max return: 93 -> 16 | Improved |
| use information concepts with TRMC instead of SAC | SAC · AntAvoid | episodes to 80% of max return: 180 -> 34 | Improved |
| use information concepts with TRMC instead of SAC | SAC · AntGather | episodes to 80% of max return: 363 -> 541 | Hurt |
| use information concepts instead of SAC policy | SAC · AntSquareWall | asymptotic performance gap: 69.0 | Hurt |
| use mutual-information metric instead of maximum likelihood metric (ablation) | TRMC - LikelihoodConcepts · AntGather | J_ap: 2095.2 -> 965.3 | Improved |
| use mutual-information metric instead of maximum likelihood metric (ablation) | TRMC - LikelihoodConcepts · AntAvoid | episodes to 50% of max return: Not reached -> 19 | Improved |
| condition the information metric on task context (ablation) | TRMC - InfoConcepts (no context) · AntGather | J_ap: 3250.5 -> 965.3 | Improved |
| condition the information metric on task context (ablation) | TRMC - InfoConcepts (no context) · AntSquareTrack | episodes to 80% of max return: Not reached -> 32 | Improved |
| condition the information metric on task context (ablation) | TRMC - InfoConcepts (no context) · AntSquareWall | episodes to 50% of max return: 25 -> 16 | Improved |
| transfer all Reptile weights instead of only fixed convolutional features (ablation) | SAC - Reptile (visual features) · AntSquareWall | J_ap: 151.4 -> 48.0 | Improved |
| use learned concepts as behavior prior on unseen maze | SAC · AntCrossMaze | convergence speed: 2 times faster to reach 85% of max return | Improved |
| transfer concepts to unseen LunarLander gravity tasks | DDQN · LunarLander | sample efficiency | Mixed |
Closest papers by meaning
- Contextual Pre-planning on Reward Machine Abstractions for Enhanced Transfer in Deep Reinforcement Learning (AAAI 2024 · similarity 0.89)
- Learning Markov State Abstractions for Deep Reinforcement Learning (NeurIPS 2021 · similarity 0.89)
- Direct then Diffuse: Incremental Unsupervised Skill Discovery for State Covering and Goal Reaching (ICLR 2022 · similarity 0.89)
- Deep Reinforcement and InfoMax Learning (NeurIPS 2020 · similarity 0.89)
- Autonomous Option Invention for Continual Hierarchical Reinforcement Learning and Planning (AAAI 2025 · similarity 0.88)
- Learning Generalizable and Composable Abstractions for Transfer in Reinforcement Learning (AAAI 2024 · similarity 0.87)
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.
