Research map / Reinforcement learning / Policy gradient and actor-critic
ACE: An Actor Ensemble Algorithm for Continuous Control with Tree Search
Shangtong Zhang, Hengshuai Yao
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 multiple independently initialized actors and select the proposal with the highest critic value instead of using a single actor (ablation) | DDPG · Roboschool tasks | number of games outperforming DDPG: 6 games (lost in 3 games and tied in 3 games) | Improved |
| add look-ahead tree search with a learned value prediction model (ablation) | Ensemble-DDPG · Roboschool tasks | number of games outperforming Ensemble-DDPG: 9 games | Improved |
| replace the value prediction model with a learned transition model (ablation) | ACE · Roboschool tasks | number of games outperforming Ensemble-DDPG: 2 games (lost in 3 games and tied in 7 games) | Hurt |
| increase network width while keeping depth comparable (ablation) | DDPG · Roboschool tasks | number of games outperforming DDPG: 1 game | No clear effect |
| share a bottom representation layer between actor and critic (ablation) | DDPG · Roboschool tasks | number of games outperforming DDPG: 2 games (lost in 2 games and tied in 8 games) | No clear effect |
| update only the selected actor at each time step rather than updating all actors (ablation) | ACE · Roboschool tasks | overall performance: ACE was slightly better overall; ACE-Alt had lower variance | Mixed |
| use a medium ensemble size (N=5) rather than the tested alternatives (ablation) | ACE · HalfCheetah | mean episode return | Improved |
| use a medium planning depth (d=1) rather than the tested alternatives (ablation) | ACE · HalfCheetah | mean episode return | Improved |
| increase the ensemble size from N=5 to N=10 at d=1 (ablation) | ACE · HalfCheetah | environment steps per second: 27.82 -> 16.12 | Hurt |
| increase planning depth from d=1 to d=2 at N=5 (ablation) | ACE · HalfCheetah | environment steps per second: 27.82 -> 6.34 | Hurt |
Closest papers by meaning
- Actor-Free Continuous Control via Structurally Maximizable Q-Functions (NeurIPS 2025 · similarity 0.87)
- Langevin Soft Actor-Critic: Efficient Exploration through Uncertainty-Driven Critic Learning (ICLR 2025 · similarity 0.86)
- On the Convergence of Continuous Single-timescale Actor-critic (ICML 2025 · similarity 0.86)
- Oracle-Efficient Reinforcement Learning for Max Value Ensembles (NeurIPS 2024 · similarity 0.86)
- Finite-Time Convergence and Sample Complexity of Actor-Critic Multi-Objective Reinforcement Learning (ICML 2024 · similarity 0.85)
- ACE: Off-Policy Actor-Critic with Causality-Aware Entropy Regularization (ICML 2024 · similarity 0.85)
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.
