Research map / Reinforcement learning / Policy gradient and actor-critic
Adversarially Guided Actor-Critic
Yannis Flet-Berliac, Johan Ferret, Olivier Pietquin, Philippe Preux, Matthieu Geist
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 an adversary-prediction discrepancy bonus without episodic count | PPO actor-critic · VizDoom | average return: 0.74 ± 0.05 vs 0. at 2M timesteps; 0.96 ± 0.001 vs 0.95 ± 0.001 at 6M timesteps | Improved |
| Add the adversarial exploration objective | PPO actor-critic · KC-S4R3 | final average performance: 0.95 vs 0.54 | Improved |
| Add the adversarial exploration objective | PPO actor-critic · KC-S5R3 | final average performance: 0.93 vs 0. | Improved |
| Add the adversarial exploration objective | PPO actor-critic · MR-N10S10 | final average performance: 0.52 vs 0.40 | Improved |
| Add the adversarial exploration objective | PPO actor-critic · OM-2Dlhb | final average performance: 0.64 vs 0.20 | Improved |
| Add the adversarial exploration objective | PPO actor-critic · OM-1Q | final average performance: 0.78 vs 0. | Improved |
| Add the adversarial exploration objective | PPO actor-critic · OM-2Q | final average performance: 0.63 vs 0. | Improved |
| Train with the adversarial objective but no extrinsic reward | AGAC · MultiRoomN10S6 | average return: ~ 0.15 vs zero average return for random agents | Improved |
| Use a smaller learning rate for the adversary than for the actor | AGAC · KeyCorridorS4R3 | average return | Improved |
| Vary the adversarial bonus coefficient c from its selected value (ablation) | AGAC · KeyCorridorS4R3 | average return | Hurt |
| Vary the actor-to-adversary learning-rate ratio ν from its selected value (ablation) | AGAC · KeyCorridorS4R3 | average return | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RIDE — RIDE: Rewarding Impact-Driven Exploration for Procedurally-Generated Environments (ICLR 2020) · better on 6 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Virtual Action Actor-Critic Framework for Exploration (Student Abstract) (AAAI 2024 · similarity 0.87)
- Studying the Interplay Between the Actor and Critic Representations in Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Value Improved Actor Critic Algorithms (NeurIPS 2025 · similarity 0.86)
- Actor-Critics Can Achieve Optimal Sample Efficiency (ICML 2025 · similarity 0.86)
- Adversarially Trained Weighted Actor-Critic for Safe Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- ReLU to the Rescue: Improve Your On-Policy Actor-Critic with Positive Advantages (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.
