1. Inicio
  2. Atrás
Título: Active learning in partially observable Markov decision processes
Autores: Jaulmes, Robin.
Fecha: 2006
Publicador: McGill University - MCGILL
Fuente:
Tipo: Electronic Thesis or Dissertation
Tema: Computer Science.
Descripción: People are efficient when they make decisions under uncertainty, even when their decisions have long-term ramifications, or when their knowledge and their perception of the environment are uncertain. We are able to experiment with the environment and learn, improving our behavior as experience is gathered. Most of the problems we face in real life are of that kind, and most of the problems that an automated agent would face in robotics too.
Our goal is to build Artificial Intelligence algorithms able to reproduce the reasoning of humans for these complex problems. We use the Reinforcement Learning framework, which allows to learn optimal behaviors in dynamic environments. More precisely, we adapt Partially-Observable Markov Decision Processes (POMDPs) to environments that are partially known.
We take inspiration from the field of Active Learning: we assume the existence of an oracle, who can, during a short learning phase, provide the agent with additional information about its environment. The agent actively learns everything that is useful in the environment, with a minimum use of the oracle.
After reviewing existing methods for solving learning problems in partially observable environments, we expose a theoretical active learning setup. We propose an algorithm, MEDUSA, and show theoretical and empirical proofs of performance for it.
Idioma: en