Tag
The paper introduces Odds-Ratio Thompson Sampling (OR-TS), a method for batched multi-armed bandits that uses contrast-based updates to handle varying common levels, showing improved regret over absolute-rate memory in simulations and real experiments.
This paper proves that Thompson sampling achieves polynomial regret for non-monotone convex ridge bandits, showing that monotonicity is not necessary, with a new regret bound of Õ(d^{9/2} √n).
Ramp Router uses EWMA for failure rates and Thompson sampling for latency to select the cheapest LLM model and service tier meeting deadlines, achieving 30% cost savings without performance loss.
This paper studies contextual causal bandits with non-manipulable variables, proposing causal variants of Thompson Sampling and Information-Directed Sampling (IDS) that exploit shared causal mechanisms to accelerate decision-making. Theoretical regret bounds and experiments on synthetic tasks show that the proposed methods outperform causal and non-causal baselines.
This paper presents a hybrid search framework that combines Thompson sampling with parallel self-avoiding walks to adaptively allocate computational effort across restriction classes for the LABS problem. The method improves previously best-known merit factors for 35 sequence lengths and achieves a new longest sequence with merit factor exceeding 8.0.
This paper proposes Absolute Thompson Sampling (ATS), a modification of Thompson Sampling that ensures optimism in expectation by using absolute exploration noise, enabling a simpler UCB-style regret analysis while maintaining computational efficiency. It achieves regret matching existing TS bounds, and introduces an ensemble variant that converges to UCB behavior.
This paper proposes a novel structured neuron pruning framework for deep neural networks using multi-armed bandit algorithms, demonstrating effectiveness on various tasks.
Purdue researchers introduce Mango, a multi-agent web navigation system that uses global site structure and Thompson Sampling to pick optimal starting URLs, boosting success rates on WebVoyager and WebWalkerQA benchmarks.