4
$\begingroup$

I am looking to use the Bayesian Bandits Strategy to find the best arm of a Multi armed bandit.

As outlined in the link, the Bayesian algorithm is

  1. Sample a random variable $X_b$ from the prior of bandit $b$, for all $b$.
  2. Select the bandit with the largest sample, i.e. select bandit $B$ = argmax $X_b$.
  3. Observe the result of pulling bandit $B$, and update the prior of bandit $B$.
  4. Return to 1.

My questions are:

  • Does each bandit have a different prior? i.e. does each bandit have an individual beta distribution, with different initial $\alpha, \beta$?
  • If each bandit does have a different prior, how do we incorporate prior knowledge about the bandits? i.e. if I suspect that bandit $1$ has a higher chance of probability than bandit $2$, how do I include this?
$\endgroup$
5
$\begingroup$

The algorithm is actually called "Thompson Sampling for Bernoulli bandits". For bandit $i$, let $X_i = 1$ if the playing is successful, and $X_i = 0$ means that playing is failure. Then, we can assume $X_i \sim Ber(p_i)$, where $p_i$ is the successful probability.

Since the conjugate prior of Bernoulli sample is Beta distribution, we assume the prior distribution $p_i\sim Beta(\alpha_i,\beta_i)$ for bandit $i$. We commonly choose the non-informative prior $\alpha_i = \beta_i = 1$ for all bandits $i=1,\cdots,K$.

Now, we can discuss about your two questions.

  • In the beginning, all bandits have the same prior distribution $Beta(1,1)$. However, after playing the game, assume the bandit $i$ receives success times $S_i$ and failure time $F_i$. Then, each bandit will have its own posterior $Beta(S_i + 1, F_i +1)$.
  • Actually, you can choose different $(\alpha_i, \beta_i)$ for each bandit $i$. Note that $\mathbb{E}(p_i) = \frac{\alpha_i}{\alpha_i + \beta_i}$ for $p_i\sim Beta(\alpha_i,\beta_i)$. If you think bandit $i$ has a higher successful probability in the very beginning, you can simply choose large $\alpha_i$ and small $beta_i$ so that the expectation of successful probability $p_i$ for bandit $i$ is very close to $1$.
$\endgroup$
1
  • $\begingroup$ Could you explain how the algorithm could be adapted to consider some context vector that would be considered at each iteration? $\endgroup$
    – guy
    Jul 7 '18 at 14:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.