Q-learning is a well-known reinforcement learning algorithm. It has
been proven mathematically that Q-learning eventually always finds
the best possible action in every state, as long as the state and action
spaces are finite. However, I will demonstrate that in practice,
Q-learning can reach very poor solutions, on some very simple
problems. (Note: convergence of Q-learning is only guaranteed for
Markov decision processes with finite state and action spaces, but the
problems I will discuss fall into this category.)

In my talk, I will attempt to give insights and answers for the following
questions:

- Why does a Q-learning agent believe that betting $1 in a game of
   roulette will on average yield $30, even after each bet has been
   attempted thousands of times?

- In what class of settings does this type of problem appear, and why?

- How does this behavior of Q-learning relate and generalize to other
   issues, such as:
 * In human decision-making, decision makers in general can seem
    overly optimistic about the choices they make.
 * In economics/finance the more you listen to your best analyst, the
    worse his/her recommendations can seem to be.
 * In a financial system, overestimation of values and underestimation of
    risks can be a result of the system and not (only) a human property
    (i.e. also in an artificial economy, with only rational agents, crises
can
    and will occur if this overestimation is not taken into account).