Research map / Bandits and online learning / Contextual bandits and Thompson sampling
Factored DRO: Factored Distributionally Robust Policies for Contextual Bandits
Tong Mu, Yash Chandak, Tatsunori Hashimoto, Emma Brunskill
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 |
|---|---|---|---|
| Separate context-shift radius (delta_x) from reward-shift radius (delta_c) instead of one joint KL radius | FDRO vs baseline-DRO · Simulated, Scene, Voting (binary rewards) | policy reward under test shift | Improved |
| Use FDRO when test context distribution is nearly identical to training (no shift) | FDRO · Simulated (Figure 1a) | reward gain over baseline-IS | Hurt |
| Increase test-population shift away from training distribution | FDRO with fixed delta_x · Simulated (Figure 1a) | reward gain over baseline-IS | Improved |
| Joint-shift DRO with binary rewards instead of factored DRO | baseline-DRO · binary-reward bandits | policy quality | Hurt |
| Increase variance regularization weight lambda in POEM | baseline-POEM · Scene and Voting (Figures 1b,c) | test reward | Hurt |
| Increase delta_x | FDRO · Scene and Voting (Figures 1b,c) | test reward | Improved |
| Set delta_x too large when no shift occurs | FDRO · Simulated (Figure 1a) | reward | Hurt |
| Sweep delta_c over a range of values (ablation) | FDRO · Scene (Figure 1f) | test reward | No clear effect |
| Exploit extra structure (factored objective) in continuous Gaussian rewards | FDRO vs baseline-DRO · Simulated continuous-reward domain (Figure 1d) | train/test performance | Improved |
| Reward-only shift environment, sweep delta_c | FDRO vs baseline-DRO · Reward-shift simulation (Figure 1e) | test reward | No clear effect |
| Per-test-population tuned delta for baselines versus fixed delta_x for FDRO | baseline-DRO, baseline-POEM · Simulated (Figure 1a) | reward gain over baseline-IS | No clear effect |
Closest papers by meaning
- Distributionally Robust Policy Evaluation and Learning in Offline Contextual Bandits (ICML 2020 · similarity 0.92)
- Offline Neural Contextual Bandits: Pessimism, Optimization and Generalization (ICLR 2022 · similarity 0.90)
- Stochastic Bandits with Context Distributions (NeurIPS 2019 · similarity 0.90)
- Practical Contextual Bandits with Regression Oracles (ICML 2018 · similarity 0.90)
- Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback (ICML 2019 · similarity 0.89)
- Model Selection for Contextual Bandits (NeurIPS 2019 · similarity 0.89)
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.
