Search sběratelské... as keywords on MyScienceWork's publications...

V posledn ch letech se uk zalo e posilovan u en je velmi slibn p i e en slo it ch her Tento v zkum se zab v aplikac technik posilov n u en na sb ratelsk karetn hry se zam en m na Hearthstone Pou ili jsme algoritmy Proximal Policy Optimization PPO a Advantage Actor-Critic A2C k tr nov n agent v r zn ch sc n ch N v zkum zahrnoval formulov n Hearthstone ve formalismu popisuj c hry s nedokonalou informac pravu simul toru hry Hearthstone v voj agent u c ch se pomoc technik posilovan ho u en definov n dostupn ch informac o stavu pro agenty implementace v ce neuronov ch s t pro akce a empirick vyhodnocen v konu agent oproti heuristick m agent m a tak proti sob V sledky uk zaly e PPO se sice dok e nau it z kladn strategie a c le hry Naproti tomu A2C vykazoval zna nou numerickou nestabilitu p i u en co jej inilo pro na e ely t m nepou iteln m Z v rem lze ci e a koli posilovac u en vykazuje ve sb ratelsk ch karetn ch hr ch potenci l sou asn algoritmy el zna n m v zv m p i dosahov n nadlidsk ho v konu v t chto multiagentn ch prost ed ch s nedokonal mi informacemi Budouc pr ce by mohla zahrnovat zkoum n sofistikovan j ch algoritm jako je RNaD a zdokonalov n metodik tr nov n s c lem zv it v konnost agent In recent years reinforcement learning has shown significant promise in solving complex games This research investigates the application of reinforcement learning techniques to Collectible Card Games CCGs with a specific focus on Hearthstone We have used Proximal Policy Optimisation PPO and Advantage Actor-Critic A2C algorithms to train agents in various scenarios Our research includes formulating Hearthstone in the formalism describing imperfect information games adjusting a Hearthstone simulator developing reinforcement learning agents defining the observation given to agents implementing multiple neural networks for actions and empirically evaluating their performance against heuristic-based agents and also against each other The results demonstrated that the PPO was able to learn the basic strategies and objectives of the game In contrast the A2C exhibited significant numerical instability rendering it nearly unusable for our purposes In conclusion while reinforcement learning shows potential in CCGs current algorithms face significant challenges in achieving superhuman performance levels in these multi-agent imperfect information environments Future work could involve exploring more sophisticated algorithms such as RNaD and refining training methodologies to enhance agent performance

Publication search with sběratelské karetní hry as keyword

Aplikace metod posilovaného učení pro sběratelské karetní hry / Applying Reinforcement Learning Techniques to Collectibl...

Publication search
with sběratelské karetní hry as keyword