Research map / Reinforcement learning / Policy gradient and actor-critic
Mind the Budget: Accelerating Deep Reinforcement Learning using Constrained Early Exit Neural Networks
Julien Brosseit, Jasper Hoffmann, Joschka Boedecker
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 budgeted early-exit actor (BEXA) to SAC/TD3 | SAC and TD3 · MuJoCo 5 tasks | Actor FLOPs speedup: up to +584% FLOPs, +95% wall-clock | Improved |
| BEXA with budget on harder tasks such as Humanoid | BEXA · MuJoCo Humanoid-v4 | Actor FLOPs speedup: +9% (SAC), +39% (TD3) | Mixed |
| Use all-exit ensemble inference instead of early-exit inference (ablation) | BEXA · MuJoCo small actors | Mean return: 62.8 -> 26.8 (SAC) | Hurt |
| Train early exits with imitation of final exit instead of DRL loss (ablation) | BEXA · MuJoCo small actors | Mean return: 62.8 -> 29.2 (SAC); 38.7 -> 34.0 (TD3) | Hurt |
| Use advantage-over-backbone gating instead of LP budget gating (ablation) | BEXA · MuJoCo small actors | Mean return: 62.8 -> 58.3 (SAC); 38.7 -> 40.5 (TD3) | Mixed |
| Use softmax-over-Q gating instead of LP budget gating (ablation) | BEXA · MuJoCo small actors | Mean return: 62.8 -> 50.3 (SAC); 38.7 -> 49.3 (TD3) | Mixed |
| Use stepwise exit training instead of joint training (ablation) | BEXA · MuJoCo small actors | Mean return: 62.8 -> 27.8 (SAC); 38.7 -> 16.5 (TD3) | Hurt |
| Greedy exit selection by Q-value without budget constraint (ablation) | early-exit actor · MuJoCo SAC training | Average exit taken | Hurt |
| Use separate critic head per exit (ablation) | BEXA · MuJoCo | learning stability | Improved |
Closest papers by meaning
- Offline-Boosted Actor-Critic: Adaptively Blending Optimal Historical Behaviors in Deep Off-Policy RL (ICML 2024 · similarity 0.87)
- ReLU to the Rescue: Improve Your On-Policy Actor-Critic with Positive Advantages (ICML 2024 · similarity 0.86)
- Overestimation, Overfitting, and Plasticity in Actor-Critic: the Bitter Lesson of Reinforcement Learning (ICML 2024 · similarity 0.86)
- The Ladder in Chaos: Improving Policy Learning by Harnessing the Parameter Evolving Path in A Low-dimensional Space (NeurIPS 2024 · similarity 0.86)
- Diffusion Actor-Critic: Formulating Constrained Policy Iteration as Diffusion Noise Regression for Offline Reinforcement Learning (ICLR 2025 · similarity 0.86)
- The Courage to Stop: Overcoming Sunk Cost Fallacy in Deep Reinforcement Learning (ICML 2025 · 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.
