Research map / Bandits and online learning / Online convex optimization and games
Meta-Learning Adversarial Bandit Algorithms
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 |
|---|---|---|---|
| Meta-learn initialization, step-size, and entropy parameter instead of fixing them task by task | Tsallis OMD for adversarial MAB · Adversarial MAB tasks with low entropy of optimal arms | task-averaged regret: O(√(sm log d)) versus Θ(√(dm)) | Helped |
| Tune the Tsallis entropy parameter over [1/log d,1] rather than fixing a single entropy (ablation) | Tsallis OMD for adversarial MAB · Implicit exploration | task-averaged regret: min_{β∈(0,1]}2√(Ĥ_β d^β m/β)+Õ((d^{3/4}m^{3/4}+d√m)/T^{1/4}) | Helped |
| Fix the entropy parameter at β=1 (Exp3) rather than tuning it (ablation) | Tsallis OMD for adversarial MAB · Implicit exploration, β=1 | task-averaged regret: 2√(Ĥ_1dm)+Õ(d^{2/3}m^{2/3}/T^{1/3}) | Mixed |
| Use guaranteed exploration and unbiased loss estimators instead of implicit exploration estimators | Tsallis OMD for adversarial MAB · MAB with a positive best-arm gap | task-averaged regret: min_{β∈(0,1]}2√(H_β d^β m/β) | Helped |
| Use guaranteed exploration without a positive best-arm gap | Tsallis OMD for adversarial MAB · MAB with unknown or small gap | expected task-averaged regret: min{min_{β∈(0,1]}2√(H_β d^β m/β)+Õ(d^{3/4}m^{1/3}/Δ),8√(dm)} | Mixed |
| Meta-learn the boundary offset and initialization | Self-concordant-barrier OMD for bandit linear optimization · Unit sphere | expected task-averaged regret: 4d√(2m log(1+(1−E||x̂̄||_2^2)/(2ε+ε^2)))+εm | Helped |
| Use a fixed task-independent initialization and boundary offset rather than meta-learning them (ablation) | Self-concordant-barrier OMD for bandit linear optimization · Identical estimated optima across tasks on the sphere | expected task-averaged regret: 1 | Hurt |
Closest papers by meaning
- Banker Online Mirror Descent: A Universal Approach for Delayed Online Bandit Learning (ICML 2023 · similarity 0.90)
- Meta-learning with Stochastic Linear Bandits (ICML 2020 · similarity 0.90)
- Provably Efficient Multi-Task Meta Bandit Learning via Shared Representations (NeurIPS 2025 · similarity 0.89)
- Online-Within-Online Meta-Learning (NeurIPS 2019 · similarity 0.89)
- Reciprocal Learning (NeurIPS 2024 · similarity 0.87)
- Adversarial Combinatorial Semi-bandits with Graph Feedback (ICML 2025 · similarity 0.87)
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.