We propose an approach for discovering in an automatic way formulas for ranking arms while playing with multi-armed bandits.
The approach works by defining a grammar made of basic elements such as for example addition, subtraction, the max operator, the average values of rewards collected by an arm, their standard deviation etc., and by exploiting this grammar to generate and test a large number of formulas. The systematic search for good candidate formulas is carried out by a built-on-purpose optimization algorithm used to navigate inside this large set of candidate formulas towards those that give high performances when using them on some multi-armed bandit problems.
We have applied this approach on a set of bandit problems made of Bernoulli, Gaussian and truncated Gaussian distributions and have identified a few simple ranking formulas that provide interesting results on every problem of this set. In particular, they clearly outperform several reference policies previously introduced in the literature. We argue that these newly found formulas as well as the procedure for generating them may suggest new directions for studying bandit problems.