Research map / Reinforcement learning / Policy gradient and actor-critic
Action Branching Architectures for Deep 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 action branches with a shared decision module instead of independent networks (ablation) | BDQ versus IDQ · MuJoCo benchmark domains | sum of rewards | Helped |
| Use action branching rather than a single joint-action output | Dueling DDQN · Reacher5DOF, n=5 | sample efficiency | Helped |
| Use action branching rather than a single joint-action output | Dueling DDQN · Reacher4DOF, n=5 | final performance | Hurt |
| Use action branching rather than a single joint-action output | Dueling DDQN · Reacher4DOF, n=9 | learning efficiency | Helped |
| Increase discretization from 17 to 33 sub-actions per dimension (ablation) | BDQ · MuJoCo benchmark domains | final performance | No clear effect |
| Replace BDQ's mean aggregation of state value and advantages with naive summation (ablation) | BDQ · continuous-control experiments | performance | Hurt |
| Replace local mean advantage reduction with local maximum reduction (ablation) | BDQ · continuous-control experiments | performance | Hurt |
| Use a mean across action branches instead of a maximum for the shared TD target (ablation) | BDQ · continuous-control experiments | performance | Helped |
| Use mean squared TD error across branches as the loss (ablation) | BDQ · continuous-control experiments | performance | Helped |
| Remove prioritized experience replay (ablation) | BDQ · MuJoCo benchmark domains | learning performance | Hurt |
| Use Gaussian exploration centered at greedy actions instead of epsilon-greedy exploration (ablation) | BDQ · standard physical control domains | performance | Helped |
| Use BDQ instead of DDPG | continuous-control agents · Humanoid-v1 | sum of rewards | Helped |
Closest papers by meaning
- Dueling Network Architectures for Deep Reinforcement Learning (ICML 2016 · similarity 0.88)
- Inferring DQN structure for high-dimensional continuous control (ICML 2020 · similarity 0.87)
- Coarse-to-fine Q-Network with Action Sequence for Data-Efficient Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Analyzing Generalization in Policy Networks: A Case Study with the Double-Integrator System (AAAI 2024 · similarity 0.84)
- SkillTree: Explainable Skill-Based Deep Reinforcement Learning for Long-Horizon Control Tasks (AAAI 2025 · similarity 0.84)
- Reinforcement learning with combinatorial actions for coupled restless bandits (ICLR 2025 · similarity 0.84)
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.