Research map / Bandits and online learning / Contextual bandits and Thompson sampling
On the Suboptimality of Thompson Sampling in High Dimensions
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 Thompson Sampling with independent uniform-prior posterior samples | Linear two-path combinatorial semi-bandit · Linear combinatorial semi-bandit, two paths | regret: R(T,θ) ≥ Δ/(4p_Δ)(1 - (1-p_Δ)^(T-1)) | Hurt |
| Use Thompson Sampling rather than a minimax-optimal strategy | Linear two-path combinatorial semi-bandit · Linear combinatorial semi-bandit, two paths, minimax | minimax regret: TS: C₄ m T^(1 - 2/d); ESCB: C₅ d^(1/4) √(T ln T) | Hurt |
| Add ℓ forced exploration rounds before Thompson Sampling (ablation) | Thompson Sampling · Linear combinatorial semi-bandit, two paths | minimax regret: C₇ C(ℓ,m) (m/ℓ) T^(1 - 1/m); C'₇ C(ℓ,m) T/ℓ | Mixed |
| Use Thompson Sampling with independent uniform-prior posterior samples | Non-linear two-decision combinatorial semi-bandit · Non-linear combinatorial semi-bandit, two decisions | regret: R(T,θ) ≥ Δ/p_Δ (1 - (1-p_Δ)^T) | Hurt |
| Use Thompson Sampling rather than CUCB | Non-linear combinatorial semi-bandit · Non-linear combinatorial semi-bandit, minimax | minimax regret: TS: C₁ T^(1 - 1/d); CUCB: C₁' d √(T ln T) | Hurt |
| Add ℓ forced exploration rounds (ablation) | Thompson Sampling on the non-linear two-decision problem · Non-linear combinatorial semi-bandit, two decisions | regret: R(T,θ) ≥ Δ/p^ℓ_Δ (1 - (1-p^ℓ_Δ)^T) | Mixed |
| Replace Thompson Sampling with ESCB and/or CUCB | Linear combinatorial semi-bandits on paths and matchings · Paths and matchings; T = 4.10^4 | regret | Helped |
| Increase the dimension m | Thompson Sampling · First selection time of the optimal decision, paths and matchings | expected first-selection time: m = 14; τ ≥ 5.10^4 with probability greater than 0.1 | Hurt |
| Decrease the gap Δ/m | Thompson Sampling · First selection time, paths and matchings | expected first-selection time | Hurt |
| Set forced exploration to ℓ = m (ablation) | Thompson Sampling on paths · First selection time of the optimal decision | expected first-selection time | No clear effect |
Closest papers by meaning
- Thompson Sampling For Combinatorial Bandits: Polynomial Regret and Mismatched Sampling Paradox (NeurIPS 2024 · similarity 0.94)
- Thompson Sampling with Less Exploration is Fast and Optimal (ICML 2023 · similarity 0.93)
- Statistical Efficiency of Thompson Sampling for Combinatorial Semi-Bandits (NeurIPS 2020 · similarity 0.92)
- Feel-Good Thompson Sampling for Contextual Bandits: a Markov Chain Monte Carlo Showdown (NeurIPS 2025 · similarity 0.90)
- The Choice of Noninformative Priors for Thompson Sampling in Multiparameter Bandit Models (AAAI 2024 · similarity 0.90)
- Geometry-Aware Approaches for Balancing Performance and Theoretical Guarantees in Linear Bandits (ICLR 2025 · 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.