Applying reinforcement learning on the bipedal walking problem

Cilj je ovog diplomskog rada predstaviti model slobodne metode učenja podrškom primjenjive na rješavanje problema neprekidne kontrole, točnije dvonožnog hodanja. U tu je svrhu prvo opisana struktura i proces učenja neuronskih mreža. Definiramo i osnovne koncepte učenja podrškom, vrste algoritama te dajemo uvid u BipedalWalker-v3 radno okruženje dvonožnog robota u kojem implementiramo algoritme. Zatim prezentiramo TD3 algoritam počevši od DDPG algoritma kao osnove istog te ARS algoritam. Konačno, opisujemo implementaciju u Pythonu i rezultate oba algoritma.The goal of this thesis is to introduce model free reinforcement learning methods applicable on solving continuous control problems, i.e. bipedal walking. For that purpose, the neural networks structure and learning process is first defined. We also define the basic concepts and algorithm types of reinforcement learning and give insight into the BipedalWalker-v3 environment in which we implement the algorithms. Subsequently, we present the TD3 algorithm (starting with the DPPG algorithm as its basis) and the ARS algorithm. Finally, we describe the Python implementation and results of both algorithms