Research map / Reinforcement learning / Policy gradient and actor-critic
Robust Reinforcement Learning with Bayesian Optimisation and Quadrature
Supratik Paul, Konstantinos Chatzilygeroudis, Kamil Ciosek, Jean-Baptiste Mouret, Michael A. Osborne, Shimon Whiteson
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 active Bayesian quadrature selection of theta instead of random sampling (ablation) | ALOQ · Robot arm approximate p(theta) | average cost: 25.95 -> 19.82 | Improved |
| use ALOQ policy over posterior instead of MAP point estimate of theta | ALOQ · Robot arm approximate p(theta) | average cost: 28.76 -> 19.82 | Improved |
| use ALOQ policy instead of RQ-ALOQ policy (ablation) | ALOQ · Robot arm approximate p(theta) | % episodes with cost >=70: 5.5% -> 0.2% | Improved |
| add intensification (second evaluation of best policy) to ALOQ (ablation) | ALOQ · Branin and Hartmann 6 | expected value of best policy: slightly underperforms one-step ALOQ | No clear effect |
| add intensification to ALOQ (ablation) | ALOQ · F-SRE1 and F-SRE2 | expected value of best policy: one-step ALOQ comparable | No clear effect |
| use ALOQ rather than WSN baseline | ALOQ · Branin | expected value of best policy: WSN substantially outperformed | Improved |
| use ALOQ rather than WSN baseline | ALOQ · Hartmann 6 | expected value of best policy: WSN better than ALOQ variants except one-step | Mixed |
| use ALOQ rather than naive and WSN | ALOQ · F-SRE1 and F-SRE2 | expected value of best policy: naive and WSN fail completely | Improved |
| use ALOQ rather than REINFORCE and TRPO | ALOQ · Robot arm collision avoidance and joint breakage | expected cost: converge to suboptimal policy even at 10000 simulator calls | Improved |
| add active selection of simulator vs physical trials (TALOQ) over ALOQ trained only in simulation | TALOQ · Physical robot ball strike | return: 4.23 -> 93.60 | Improved |
| use TALOQ rather than ALOQ trained only on the proxy for reality | TALOQ · Simulated robot arm (proxy for reality) | sample efficiency | Improved |
| vary threshold k in TALOQ (ablation) | TALOQ · Simulated robot arm | learning speed | No clear effect |
Closest papers by meaning
- Alternating Optimisation and Quadrature for Robust Control (AAAI 2018 · similarity 0.98)
- Distributionally Robust $Q$-Learning (ICML 2022 · similarity 0.91)
- Minimax Optimal and Computationally Efficient Algorithms for Distributionally Robust Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- Minimax Optimal Reinforcement Learning with Quasi-Optimism (ICLR 2025 · similarity 0.88)
- Bayesian Exploration Networks (ICML 2024 · similarity 0.88)
- Distributionally Robust Reinforcement Learning with Interactive Data Collection: Fundamental Hardness and Near-Optimal Algorithms (NeurIPS 2024 · similarity 0.88)
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.
