Modèles de monde d'apprentissage pour les jeux de type puzzle
Théo Michel, Levit, Svetlana · HAL (Le Centre pour la Communication Scientifique Directe) · 2023
This work was part of the course Learning and Intelligent Systems during the winter semester 2022-2023 at the TU Berlin. Recently, we have seen a rise of model-based algorithms in reinforcement learning, whose main advantage lies in the training efficiency enabled by representing the world in a compressed latent representation. One such algorithm is Dreamer V2, which has surpassed human-level performance on all Atari games, and has also been used to teach a quadruped to walk using only real-world experience [7]. However, we have not seen the usefulness of the Dreamer V2 algorithm when applied to more complicated logical scenarios, and we cannot really predict the results because we do not really have a deep understanding of the world model. The following is a summary article of the exploration of the Dreamer V2 algorithm applied to complex tasks. We will examine the results of the algorithm when applied to different tasks (puzzles, exploration game...) with different levels of difficulty and logical complexity. And we will also examine the quality of the generated world models, using simple methods to analyse input and output correlations, and laying the groundwork for more advanced world model analysis.