In this talk, I will provide a short overview of recent results in multi-armed bandit for reinforcement learning theory.

I will show how a novel paradigm from revent advances in bandit theory is reshaping the exploration-exploitation challenge at large, yielding improved algorithms  including in full-blown reinforcement learning.

We will also explore the strucutre of generic MDPs.

In the second part, I will discuss a stimulating example of the sim-to-real gap in experimental sciences (especially agroecosystems), revealing a fresh set of challenges to be adressed both from an applied and methodological standpoint.