@Article{j-Urteaga2018,
  author        = {I{\~n}igo Urteaga and Chris H. Wiggins},
  journal       = {arXiv e-print:1808.02933},
  title         = {{(Sequential) Importance Sampling Bandits}},
  year          = {2018},
  month         = aug,
  abstract      = {This work extends existing multi-armed bandit (MAB) algorithms beyond their original settings by leveraging advances in sequential Monte Carlo (SMC) methods from the approximate inference community. We leverage Monte Carlo estimation and, in particular, the flexibility of (sequential) importance sampling to allow for accurate estimation of the statistics of interest within the MAB problem. The MAB is a sequential allocation task where the goal is to learn a policy that maximizes long term payoff, where only the reward of the executed action is observed; i.e., sequential optimal decisions are made, while simultaneously learning how the world operates. In the stochastic setting, the reward for each action is generated from an unknown distribution. To decide the next optimal action to take, one must compute sufficient statistics of this unknown reward distribution, e.g., upper-confidence bounds (UCB), or expectations in Thompson sampling. Closed-form expressions for these statistics of interest are analytically intractable except for simple cases. By combining SMC methods --- which estimate posterior densities and expectations in probabilistic models that are analytically intractable --- with Bayesian state-of-the-art MAB algorithms, we extend their applicability to complex models: those for which sampling may be performed even if analytic computation of summary statistics is infeasible --- nonlinear reward functions and dynamic bandits. We combine SMC both for Thompson sampling and upper confident bound-based (Bayes-UCB) policies, and study different bandit models: classic Bernoulli and Gaussian distributed cases, as well as dynamic and context dependent linear-Gaussian, logistic and categorical-softmax rewards.},
  archiveprefix = {arXiv},
  eprint        = {1808.02933},
  file          = {:http\://arxiv.org/pdf/1808.02933v3:PDF},
  keywords      = {stat.ML, cs.LG, stat.CO, I.2.6},
  primaryclass  = {stat.ML},
}

{"_id":"PRMrx4rc9TGywx9xS","bibbaseid":"urteaga-wiggins-sequentialimportancesamplingbandits-2018","author_short":["Urteaga, I.","Wiggins, C. H."],"bibdata":{"bibtype":"article","type":"article","author":[{"firstnames":["Iñigo"],"propositions":[],"lastnames":["Urteaga"],"suffixes":[]},{"firstnames":["Chris","H."],"propositions":[],"lastnames":["Wiggins"],"suffixes":[]}],"journal":"arXiv e-print:1808.02933","title":"(Sequential) Importance Sampling Bandits","year":"2018","month":"August","abstract":"This work extends existing multi-armed bandit (MAB) algorithms beyond their original settings by leveraging advances in sequential Monte Carlo (SMC) methods from the approximate inference community. We leverage Monte Carlo estimation and, in particular, the flexibility of (sequential) importance sampling to allow for accurate estimation of the statistics of interest within the MAB problem. The MAB is a sequential allocation task where the goal is to learn a policy that maximizes long term payoff, where only the reward of the executed action is observed; i.e., sequential optimal decisions are made, while simultaneously learning how the world operates. In the stochastic setting, the reward for each action is generated from an unknown distribution. To decide the next optimal action to take, one must compute sufficient statistics of this unknown reward distribution, e.g., upper-confidence bounds (UCB), or expectations in Thompson sampling. Closed-form expressions for these statistics of interest are analytically intractable except for simple cases. By combining SMC methods — which estimate posterior densities and expectations in probabilistic models that are analytically intractable — with Bayesian state-of-the-art MAB algorithms, we extend their applicability to complex models: those for which sampling may be performed even if analytic computation of summary statistics is infeasible — nonlinear reward functions and dynamic bandits. We combine SMC both for Thompson sampling and upper confident bound-based (Bayes-UCB) policies, and study different bandit models: classic Bernoulli and Gaussian distributed cases, as well as dynamic and context dependent linear-Gaussian, logistic and categorical-softmax rewards.","archiveprefix":"arXiv","eprint":"1808.02933","file":":http\\://arxiv.org/pdf/1808.02933v3:PDF","keywords":"stat.ML, cs.LG, stat.CO, I.2.6","primaryclass":"stat.ML","bibtex":"@Article{j-Urteaga2018,\n author = {I{\\~n}igo Urteaga and Chris H. Wiggins},\n journal = {arXiv e-print:1808.02933},\n title = {{(Sequential) Importance Sampling Bandits}},\n year = {2018},\n month = aug,\n abstract = {This work extends existing multi-armed bandit (MAB) algorithms beyond their original settings by leveraging advances in sequential Monte Carlo (SMC) methods from the approximate inference community. We leverage Monte Carlo estimation and, in particular, the flexibility of (sequential) importance sampling to allow for accurate estimation of the statistics of interest within the MAB problem. The MAB is a sequential allocation task where the goal is to learn a policy that maximizes long term payoff, where only the reward of the executed action is observed; i.e., sequential optimal decisions are made, while simultaneously learning how the world operates. In the stochastic setting, the reward for each action is generated from an unknown distribution. To decide the next optimal action to take, one must compute sufficient statistics of this unknown reward distribution, e.g., upper-confidence bounds (UCB), or expectations in Thompson sampling. Closed-form expressions for these statistics of interest are analytically intractable except for simple cases. By combining SMC methods --- which estimate posterior densities and expectations in probabilistic models that are analytically intractable --- with Bayesian state-of-the-art MAB algorithms, we extend their applicability to complex models: those for which sampling may be performed even if analytic computation of summary statistics is infeasible --- nonlinear reward functions and dynamic bandits. We combine SMC both for Thompson sampling and upper confident bound-based (Bayes-UCB) policies, and study different bandit models: classic Bernoulli and Gaussian distributed cases, as well as dynamic and context dependent linear-Gaussian, logistic and categorical-softmax rewards.},\n archiveprefix = {arXiv},\n eprint = {1808.02933},\n file = {:http\\://arxiv.org/pdf/1808.02933v3:PDF},\n keywords = {stat.ML, cs.LG, stat.CO, I.2.6},\n primaryclass = {stat.ML},\n}\n\n","author_short":["Urteaga, I.","Wiggins, C. H."],"key":"j-Urteaga2018","id":"j-Urteaga2018","bibbaseid":"urteaga-wiggins-sequentialimportancesamplingbandits-2018","role":"author","urls":{},"keyword":["stat.ML","cs.LG","stat.CO","I.2.6"],"metadata":{"authorlinks":{}}},"bibtype":"article","biburl":"https://iurteaga.github.io/myJournals.bib","dataSources":["DaNNNZA5viNZwqGNn"],"keywords":["stat.ml","cs.lg","stat.co","i.2.6"],"search_terms":["sequential","importance","sampling","bandits","urteaga","wiggins"],"title":"(Sequential) Importance Sampling Bandits","year":2018}