Building perfectly curious machines: using structural causal modeling to define the ideal feature space at the learning baseline of curiosity-driven agents

The thesis develops an ideal inverse-dynamics learning algorithm which can learn the properties of the sensors and actuators under its control. The algorithm converges on an ideal feature space, where the implementation details of the actuators under the algorithm's control are rendered invisible to the forward dynamics of a curiosity-driven algorithm (with the same sensors and actuators), run on top of that feature space, where the curiosity-driven algorithm's reward is strictly determined by the minimization of the error of his prediction of the next state of his environment given the current state and his action. That is, the ideal feature space allows the learning trajectory of the forward dynamics of a curiosity-driven algorithm to concentrate on the dynamics of the algorithm's environment by avoiding any distractions originating in the properties of the sensors and actuators under the algorithm's control.

• Autores: Orozco García, Tomás
• Fecha de publicación: 2020
• Institución: Universidad de los Andes
• Repositorio: Séneca: repositorio Uniandes
• Idioma: eng
• OAI Identifier: oai:repositorio.uniandes.edu.co:1992/50981
• Acceso en línea: http://hdl.handle.net/1992/50981
• Palabra clave: Aprendizaje por refuerzo (Aprendizaje automático); Aprendizaje automático (Inteligencia artificial); Aprendizaje impulsado por la curiosidad; Algoritmos (Computadores); Ingeniería
• Rights: openAccess
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/