Meta-learning for deep reinforcement learning

• 主要作者: Poon, Jun Yaw
• 其他作者: Sinno Jialin Pan
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2021
• 主題: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• 在線閱讀: https://hdl.handle.net/10356/148139
• 機構: Nanyang Technological University

Creating a fully autonomous agent that is able to solve real world tasks at superhuman level is always one of the primary goals in the field of artificial intelligence (AI). Most works that are done in AI field until today are categorized as ”narrow” AI, even though the field of AI has been in the world for seventy years [1]. A more ambitious goal in AI is to create Artificial General Intelligence (AGI). The concept of meta-learning is one step closer to this ambitious goal of AI. In this project, we studied existing meta-learning algorithms for deep reinforcement learning, namely MAML and PEARL, and compare them in terms of sample efficiency and their performance in terms of the final average returns. We also studied an existing algorithm in RL, namely SLAC, that addresses the latent representation problem in RL by learning a sequential latent variable model. We integrated SLAC into PEARL and compare the its performance to the original PEARL. Our work shows that the integrated variant of PEARL with SLAC has a better performance than the original PEARL.