Teaching a neural network to play a simple arcade game

V diplomskem delu smo izdelali nevronsko mrežo in preprosto arkadno videoigro Bomberman, katero se je nevronska mreža naučila igrati. Videoigro smo izdelali s pomočjo odprtokodnega ogrodja LibGdx v programskem jeziku Java. Implementirali smo lastno nevronsko mrežo in nevronsko mrežo s knjižnico PyTorch. Obe mreži se učita po principu okrepitvenega učenja. Rezultate lastne implementacije nevronske mreže smo primerjali z rezultati nevronske mreže implementirane s knjižnico PyTorch. Ugotovili smo, da so si rezultati obeh mrež zelo podobni. Nevronska mreža po tridesetih minutah učenja uspe počistiti polovico vseh uničljivih zidov.The thesis aimed to create a neural network and a simple arcade video game called Bomberman that the neural network would learn to play. The video game was created using the open-source framework LibGdx in the Java programming language. We implemented our own neural network and a neural network based on the PyTorch library. Both networks learn using the reinforcement learning method. The results of our own implementation of the neural network were compared with the results of the neural network implemented with the PyTorch library. We found that the results of both networks are very similar. After thirty minutes of learning, the neural network manages to clear half of all destructible walls