Research map / Bandits and online learning / Multi-armed bandits
Regret Minimisation in Multi-Armed Bandits Using Bounded Arm Memory
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 |
|---|---|---|---|
| Use bounded-memory phased exploration with UCB-M on infinite Bernoulli arms instead of non-stationary Policy | QUCB-M (M=2) · I1–I4 | cumulative regret (/10^5): 3.58 ±0.4 -> 1.84±0.17; 1.11±0.2 -> 0.41±0.02; 1.64 ± 0.2 -> 1.29±0.10; 0.79 ± 0.1 -> 0.49±0.02 | Helped |
| Use bounded-memory phased exploration with UCB-M instead of non-recalling √T-run | QUCB-M (M=2) · I1, I3, I4 | cumulative regret (/10^5): 5.35 ±0.5 -> 1.84±0.17; 4.56 ± 0.001 -> 1.29±0.10; 2.55 ± 0.001 -> 0.49±0.02 | Helped |
| Use MOSS instead of UCB allocation (ablation) | Quantile-regret algorithm · I1–I4, M=2 | cumulative regret (/10^5): 1.84±0.17 -> 1.74±0.17; 0.41±0.02 -> 0.31±0.02; 1.29±0.10 -> 1.20±0.10; 0.49±0.02 -> 0.39±0.02 | Helped |
| Use Thompson Sampling instead of UCB allocation (ablation) | Quantile-regret algorithm · I1–I4, M=2 | cumulative regret (/10^5): 1.84±0.17 -> 1.77±0.17; 0.41±0.02 -> 0.32±0.04; 1.29±0.10 -> 1.23±0.09; 0.49±0.02 -> 0.40±0.02 | Mixed |
| Set UCB confidence multiplier η=0.2 (ablation) | QUCB-M · I1–I4, M=2 | cumulative regret (/10^5): 1.84±0.17 -> 2.00±0.20; 0.41±0.02 -> 0.32±0.05; 1.29±0.10 -> 1.41±0.10; 0.49±0.02 -> 0.69±0.04 | Mixed |
| Increase memory from M=2 to M=10 (ablation) | QUCB-M (η=1) · I1–I4 | cumulative regret (/10^5): 1.84±0.17 -> 1.98±0.16; 0.41±0.02 -> 0.59±0.02; 1.29±0.10 -> 1.49±0.09; 0.49±0.02 -> 0.63±0.01 | Hurt |
| Increase memory from M=2 to M=10 (ablation) | QUCB-M (η=0.2) · I1–I4 | cumulative regret (/10^5): 2.00±0.20 -> 1.71±0.16; 0.32±0.05 -> 0.16±0.02; 1.41±0.10 -> 1.16±0.09; 0.69±0.04 -> {0.30}±0.02 | Helped |
| Use M=10 rather than M=2 (ablation) | QTS-M · I1–I4 | cumulative regret (/10^5): 1.77±0.17 -> 1.91±0.16; 0.32±0.04 -> 0.18±0.03; 1.23±0.09 -> 1.14±0.10; 0.40±0.02 -> {0.30}±0.02 | Mixed |
| Use M=10 rather than M=2 (ablation) | QMoss-M · I1–I4 | cumulative regret (/10^5): 1.74±0.17 -> 1.69±0.15; 0.31±0.02 -> 0.25±0.02; 1.20±0.10 -> 1.13±0.09; 0.39±0.02 -> {0.30}±0.010 | Mixed |
| Use α=0.205 rather than α=0.347 for arm exploration (ablation) | QRMUCBM · Infinite Bernoulli instances | cumulative regret | Hurt |
Closest papers by meaning
- Multi-Armed Bandits with Bounded Arm-Memory: Near-Optimal Guarantees for Best-Arm Identification and Regret Minimization (NeurIPS 2021 · similarity 0.95)
- Tight Regret Bounds for Single-pass Streaming Multi-armed Bandits (ICML 2023 · similarity 0.93)
- Bandits with many optimal arms (NeurIPS 2021 · similarity 0.92)
- Improved Regret Bounds for Linear Bandits with Heavy-Tailed Rewards (NeurIPS 2025 · similarity 0.91)
- p-Mean Regret for Stochastic Bandits (AAAI 2025 · similarity 0.91)
- Bayesian Regret Minimization in Offline Bandits (ICML 2024 · similarity 0.91)
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.