Representation learning for hierarchical reinforcement learning

• Autores: Lorenzo Steccanella
• Directores de la Tesis: Anders Jonsson (dir. tes.)
• Lectura: En la Universitat Pompeu Fabra ( España ) en 2023
• Idioma: español
• Tribunal Calificador de la Tesis: Vicente Gómez Cerdà (presid.), Sergio Jiménez Celorrio (secret.), Sindri Magnússon (voc.)
• Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y las Comunicaciones por la Universidad Pompeu Fabra
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Inteligencia artificial
• Enlaces
  • Tesis en acceso abierto en: TDX
• Resumen
  • español

    El método Hiearchical Reinforcement Learning (HRL) tiene el potencial de simplificar la solución de entornos con escasas recompensas y horizontes a largo plazo. La idea detrás de HRL es descomponer un problema complejo de toma de decisiones en subproblemas más pequeños y manejables, lo que permite que un agente aprenda de manera más eficiente y efectiva. En esta tesis, pretendemos contribuir al campo del HRL a través del estudio de las representaciones de partición del espacio de estado. Nuestro objetivo es descubrir representaciones que permitan descomponer un espacio de estado complejo en un conjunto de particiones interconectadas. Comenzamos nuestro trabajo presentando cuáles son las propiedades de las particiones de espacio de estado ideales para HRL y luego procedemos a explorar diferentes métodos para crear dichas particiones. Presentamos algoritmos capaces de aprovechar tales representaciones para aprender de manera más efectiva en entornos de escasa recompensa. Finalmente, mostramos cómo combinar la representación aprendida con el método Goal-Conditioned Reinforcement Learning (GCRL) y, adicionalmente, presentamos representaciones de estado útiles para GCRL.

  • English

    The concept of hierarchy has a strong appeal to researchers in the field of artificial intelligence. Humans cope with the complexity of the world by thinking in a hierarchical manner, and this same faculty is sought to be imparted to autonomous agents.

    Over the past few years, Reinforcement Learning (RL) methods have achieved remarkable success, largely facilitated by the use of deep learning models, reaching human-level performance in several domains, including Atari games and complex board games such as Go and Chess. Nevertheless, it is still a challenge for RL agents to solve environments with sparse rewards and long time horizons.

    Hierarchical Reinforcement Learning (HRL) has the potential to simplify the solution of such environments. The idea behind HRL is to decompose a complex decision-making problem into smaller, manageable sub-problems, allowing an agent to learn more efficiently and effectively.

    In this thesis, we aim to contribute to the field of HRL through the study of state space partition representations. We aim to discover representations that allow decomposing a complex state space in a set of small interconnected partitions. We start our work by presenting which are the properties of ideal state space partitions for HRL and then proceed to explore different methods for creating such partitions. We present algorithms able to leverage such representations to learn more effectively in sparse reward settings. Finally, we show how to combine the learned representation with Goal-Conditioned Reinforcement Learning (GCRL) and additionally we present state representations useful for GCRL.