Research map / Reinforcement learning / Visual and representation RL
Latent Space Policies for Hierarchical Reinforcement Learning
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 |
|---|---|---|---|
| Use SAC-LSP as a single-level policy instead of the tested prior RL methods | SAC-LSP · OpenAI Gym continuous-control benchmarks | expected return | Mixed |
| Add a second latent-policy layer, trained stagewise on top of a frozen base policy | SAC-LSP · Ant | performance | Helped |
| Add a second latent-policy layer, trained stagewise on top of a frozen base policy | SAC-LSP · Humanoid | performance | Helped |
| Train a four-layer invertible policy end-to-end instead of adding a second layer stagewise (ablation) | SAC-LSP · Ant | performance | No clear effect |
| Train a four-layer invertible policy end-to-end instead of adding a second layer stagewise (ablation) | SAC-LSP · Humanoid | performance | Hurt |
| Pretrain a lower-level policy with a motion reward, then train a higher-level policy on the task reward | SAC-LSP · Ant maze navigation | learning speed | Helped |
| Train a single policy from scratch using only the sparse task reward (ablation) | Single-level policy · Ant maze navigation | task success | Hurt |
| Fine-tune a pretrained locomotion policy using only the task reward (ablation) | Single-level policy · Ant maze navigation | learning speed | Helped |
| Add a motion shaping reward to the entire policy while training from scratch (ablation) | Single-level policy · Ant maze navigation | learning speed | Helped |
| Add a motion shaping reward to the entire policy while fine-tuning a pretrained policy (ablation) | Single-level policy · Ant maze navigation | learning speed | Helped |
| Use motion shaping for the full policy instead of training the higher layer on the task reward alone (ablation) | Single-level policy · Ant maze navigation | final distance to goal | Hurt |
| Use latent-space policy learning with SAC instead of applying the method to soft Q-learning (ablation) | Latent-space policy learning · Ant maze navigation | learning speed | Helped |
Closest papers by meaning
- Stochastic Neural Networks for Hierarchical Reinforcement Learning (ICLR 2017 · similarity 0.91)
- Sub-policy Adaptation for Hierarchical Reinforcement Learning (ICLR 2020 · similarity 0.90)
- Hierarchical Reinforcement Learning via Advantage-Weighted Information Maximization (ICLR 2019 · similarity 0.89)
- Learning Subgoal Representations with Slow Dynamics (ICLR 2021 · similarity 0.88)
- Improving Unsupervised Hierarchical Representation with Reinforcement Learning (CVPR 2024 · similarity 0.88)
- Probabilistic Subgoal Representations for Hierarchical 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.