Bootstrapped Thompson Sampling and Deep Exploration

Abstract

This technical note presents a new approach to carrying out the kind of exploration achieved by Thompson sampling, but without explicitly maintaining or sampling from posterior distributions. The approach is based on a bootstrap technique that uses a combination of observed and artificially generated data. The latter serves to induce a prior distribution which, as we will demonstrate, is critical to effective exploration. We explain how the approach can be applied to multi-armed bandit and reinforcement learning problems and how it relates to Thompson sampling. The approach is particularly well-suited for contexts in which exploration is coupled with deep learning, since in these settings, maintaining or generating samples from a posterior distribution becomes computationally infeasible.

References in corpus (3)

• Further Optimal Regret Bounds for Thompson Sampling
• Model-based Reinforcement Learning and the Eluder Dimension
• Thompson sampling with the online bootstrap

Cited by in corpus (7)

• Variational Deep Q Network
• Risk Aware and Multi-Objective Decision Making with Distributional Monte Carlo Tree Search
• Residual Bootstrap Exploration for Bandit Algorithms
• Differentiable Linear Bandit Algorithm
• CORe: Capitalizing On Rewards in Bandit Exploration
• State-Aware Variational Thompson Sampling for Deep Q-Networks
• Knowledge is reward: Learning optimal exploration by predictive reward cashing