Learning to park a car in a simulator using DDPG algorithm

Diplomska naloga raziskuje problem parkiranja avtomobila v simulatorju s pomoˇcjo algoritma spodbujevanega uˇcenja DDPG. V nalogi se spoznamo s teoretiˇcno podlago spodbujevanega uˇcenja in nevronskih mreˇz ter si bolj podobno pogledamo algoritem DDPG. Glede na pridobljeno znanje implementiramo agenta, ki parkira na praznem parkiriˇsˇcu. Primerjamo, kako se razliˇcne arhitekture nevronske mreˇze obnesejo na problemu in kako globina in ˇsirina mreˇze vplivata na rezultate. Primerjamo jih na podlagi odstotka uspeˇsnih parkiranj, povpreˇcnega ˇstevila korakov za uspeˇsno parkiranje in poti, ki jih avtomobil opravi med parkiranjem. Najbolj uspeˇsna arhitektura je problem parkiranja in nakljuˇcne toˇcke reˇsila 100-odstotno v povpreˇcno 20 korakih. To arhitekturo smo testiral ˇse na poligonih z ovirami, ki so predstavljali postopno teˇzje oblike ˇcelnega, vzvratnega in boˇcnega parkiranja. Rezultati so obetavni in ponujajo moˇznost za nadaljnje raziskovanje.The thesis explores the problem of parking inside a simulator with the help of a reinforcement learning algorithm DDPG. We get familiar with the theoretical background of reinforcement learning, neural networks, and an in-depth knowledge of DDPG. Based on our knowledge we implement an agent capable of parking in an empty parking lot. We compare different neural network architectures and how changing the depth and width affect the results. We compare the results based on the percentage of successful episodes, the average steps necessary for a successful episode, and the paths the car made during parking. The most successful architecture solved the problem of parking from a random starting point 100% and in on average 20 steps. We then tested this architecture on courses with obstacles that represented gradually harder degrees of difficulty for perpendicular, reverse and parallel parking. The results are promising and offer room for further research and developmen