Thompson Sampling
A Bayesian inference-based contextual bandit algorithm that naturally incorporates exploration uncertainty by sampling from the posterior reward distribution of each option. It is widely applied in personalized recommendation and online decision-making scenarios.
Timeline (last 90 days)
TS² (Thompson Sampling squared) is a sampling-based variant of Thompson sampling, and the 'squared' in its name hints at the central role of the squared confidence-width term in the sampling regret.
Common algorithmic implementations of contextual bandits include LinUCB (Linear Upper Confidence Bound) and Thompson Sampling.
Thompson Sampling takes a Bayesian perspective, performing posterior sampling on the reward distribution of each option to incorporate uncertainty into the exploration strategy.
LinUCB and Thompson Sampling generally perform more stably in practice than a fixed ε-greedy strategy.
At the system level, exploration mechanisms (such as Epsilon-Greedy or Thompson Sampling) can be designed to allocate exposure to new entities in order to address the cold-start problem.
All Facts (5)
TS² (Thompson Sampling squared) is a sampling-based variant of Thompson sampling, and the 'squared' in its name hints at the central role of the squared confidence-width term in the sampling regret.
50%UnverifiedCommon algorithmic implementations of contextual bandits include LinUCB (Linear Upper Confidence Bound) and Thompson Sampling.
50%UnverifiedThompson Sampling takes a Bayesian perspective, performing posterior sampling on the reward distribution of each option to incorporate uncertainty into the exploration strategy.
50%UnverifiedLinUCB and Thompson Sampling generally perform more stably in practice than a fixed ε-greedy strategy.
50%UnverifiedAt the system level, exploration mechanisms (such as Epsilon-Greedy or Thompson Sampling) can be designed to allocate exposure to new entities in order to address the cold-start problem.
50%