This work aims to analyze the performance of an agent based on Reinforcement Learning. Your learning engine is based on three algorithms: Qlearning (QL), Deep Q-Network (DQN) and Double Deep Q-Network (DDQN). To validate the agent and its methods, it was defined as environment the World of Wumpus, which was modeled according to the environment standards adopted by DeepMind Lab. From the experiments performed and their respective configurations, it was observed that the agents managed to reach the main objective only in two configurations of environments. In the 4x4 environment the winning percentage of the QL, DQN algorithms and DDQN were 0.005, 22.96, 18.73% respectively, which drastically reduced specifically for the 10x10 scenario and failing to meet the objective for the other environments.
|
