Usando Aprendizado por Reforçoo com Demonstrações e Habilidades para Jogar Sonic com Recompensas Esparsas

Abstract

Reinforcement learning can handle many types of environments, but it does so inefficiently, requiring millions of learning steps to achieve good per formance, which limits its use in the real world. Using demonstrations on pre training and skills, a sequence of actions that perform a sub-task, we trained an agent that can complete the first level and achieve progress in the second and third levels, which the agent has never seen before, of the game Sonic The Hedgehog. We used sparse rewards, with rewards being given only when a level is completed, something that common reinforcement learning techniques would not be able to do, andimageinput using Skill-based Model-based Reinforcement Learning, originally used in robotics. Our results show how the use of demons trations and skills can speed up training and enable the use of reinforcement learning in a sparse reward environment.