Thompson sampling - NISHIO Hirokazu's Scrapbox (Auto-translated from Japanese)

Thompson sampling

Theory and Algorithms for the Bandit Problem p.38 Thompson extraction

Bayesian estimation] of expected value

Choose the action with the probability of being the maximum expected value of each action (random dither)

However, instead of doing this "probability of being the maximum expected value" calculation, use the [random-choice algorithm

Since it is Bayesian, a distribution of expected values is obtained. Sampling from this distribution

Select the action that had the largest value as a result of sampling

This will make it possible to "choose that action with the probability of being the maximum expected value.

https://hagino3000.blogspot.com/2015/07/thompson-sampling.html

https://hagino3000.blogspot.com/2016/12/linear-bandit.html

#Reinforcement Learning

---

This page is auto-translated from /nishio/トンプソンサンプリング using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.