2017/06/14 by Adam N. Elmachtoub, Elmachtoub, Adam N., Ryan McNellis +5 · 2 citations
Computer Science · Decision Sciences · #Advanced Bandit Algorithms Research #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Algorithms #Reinforcement Learning in Robotics
paper · pdf · doi:10.48550/arxiv.1706.04687
openalex publication_date 2017/06/14 · openalex created_date 2022/09/30 · openalex updated_date 2026/07/28
Many efficient algorithms with strong theoretical guarantees have been\nproposed for the contextual multi-armed bandit problem. However, applying these\nalgorithms in practice can be difficult because they require domain expertise\nto build appropriate features and to tune their parameters. We propose a new\nmethod for the contextual bandit problem that is simple, practical, and can be\napplied with little or no domain expertise. Our algorithm relies on decision\ntrees to model the context-reward relationship. Decision trees are\nnon-parametric, interpretable, and work well without hand-crafted features. To\nguide the exploration-exploitation trade-off, we use a bootstrapping approach\nwhich abstracts Thompson sampling to non-Bayesian settings. We also discuss\nseveral computational heuristics and demonstrate the performance of our method\non several datasets.\n