Título: Risk-directed exploration in reinforcement learning
Autores: Law, Edith L. M.
Fecha: 2005
Publicador: McGill University - MCGILL
Fuente:
Tipo: Electronic Thesis or Dissertation
Tema: Computer Science.
Descripción: Reinforcement Learning is a class of methods for solving sequential decision problems when the model of the environment is not known. In this framework, the agent must explore the environment to gather more information about the model and the utility of each of its actions, while striving to act as well as possible using limited knowledge. One of the major obstacles that prevent reinforcement learning from being extended to real-life settings is the fact that the agent is blind to the risk of actions during learning, potentially ending up in catastrophic states. This thesis presents a model-based directed exploration method for selecting actions based on a measure of risk, characterized by entropy and expected immediate reward. The weighted combination of this risk measure and the long term utility of the action, or risk-adjusted utility, is used to determine the probability of different actions. Using this approach, agents can manifest risk-averse or risk-seeking behavior. Experimental results show that risk-directed exploration can result in better performance during learning than the standard Boltzmann action selection method, or other directed exploration methods such as counter-based and recency-based methods.
Idioma: en