Basics of reinforcement learning

Ovaj se rad bavi osnovama podržanog učenja iznoseći sve poznate metode korištene u PU poput Bellmanovih jednadžbi, Dinamičkog programiranja, teorema poboljšanja strategije,Monte-Carlo učenja, učenja s vremenskom razlikom itd. Osnovni problem je bio nalaženje optimalne strategije uz pomoć optimalne vrijednosne funkcije, i to u okolini bez modela. Kombinacijom navedenih metoda, posebno MC iteracije strategija, napisan je kod koji je uspješno nalazio optimalne strategije na primjeru igre Blackjack.This paper deals with the basics of reinforcement learning, presenting all known methods used in PU such as Bellman equations, Dynamic programming, strategy improvement theorem, MC learning, temporal difference learning, etc. The basic problem was finding an optimal strategy that converges to an optimal value function, and in a model-free environment. By combining the above methods, especially MC strategy iteration, a code was written that successfully found optimal strategies on the example of the Blackjack game