Research map / Reinforcement learning / MDPs and planning
Adaptive Policy Backbone via Shared Network
Bumgeun Park, Donghwan Lee
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 |
|---|---|---|---|
| freeze a randomly initialized backbone and train only pre/post linear layers instead of end-to-end RL | APB (random backbone) · Cheetah-vel and Cheetah-dir | average return | No clear effect |
| freeze a randomly initialized backbone and train only linear layers | APB (random backbone) · Ant-goal | average return | Improved |
| freeze a randomly initialized backbone and train only linear layers | APB (random backbone) · Ant-dir | average return | Hurt |
| adapt both pre- and post-backbone linear layers instead of post-backbone only (linear probing) (ablation) | APB (random backbone) · MuJoCo meta-RL tasks | average return | Improved |
| use APB with meta-trained frozen backbone instead of fine-tuning-based or context-based meta-RL baselines | APB · OOD adaptation (Cheetah-vel, Cheetah-vel to Cheetah-dir, Ant-goal, Ant-dir) | average return | Improved |
| use meta-trained frozen backbone instead of randomly initialized frozen backbone in behavior cloning (ablation) | APB-BC · BC on OOD tasks with narrow demonstrations | mean episodic return | Improved |
| freeze a randomly initialized backbone instead of training the full model in behavior cloning (ablation) | Random APB-BC · BC on OOD tasks with narrow demonstrations | mean episodic return | Improved |
| use meta-trained versus random frozen backbone in online RL with extended episode length (ablation) | APB · Extended episode length (H_eval > H_train) | average return | No clear effect |
| freeze backbone and train only linear layers (parameter efficiency) | APB · Cheetah-vel | trainable parameters and time per iteration: 219,606 -> 6,906; 5.65 ms -> 4.78 ms | Improved |
| freeze backbone and train only linear layers (parameter efficiency) | APB · Ant-goal | trainable parameters and time per iteration: 247,708 -> 35,008; 5.61 ms -> 4.92 ms | Improved |
Closest papers by meaning
- Provably Efficient Model-based Policy Adaptation (ICML 2020 · similarity 0.88)
- Scalable Multi-Objective and Meta Reinforcement Learning via Gradient Estimation (AAAI 2026 · similarity 0.88)
- The Ladder in Chaos: Improving Policy Learning by Harnessing the Parameter Evolving Path in A Low-dimensional Space (NeurIPS 2024 · similarity 0.88)
- Meta-Reinforcement Learning with Universal Policy Adaptation: Provable Near-Optimality under All-task Optimum Comparator (NeurIPS 2024 · similarity 0.87)
- PWM: Policy Learning with Multi-Task World Models (ICLR 2025 · similarity 0.87)
- Efficient Recurrent Off-Policy RL Requires a Context-Encoder-Specific Learning Rate (NeurIPS 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.
