A tutorial survey of reinforcement learning (Q5955768): Difference between revisions

Reinforcement learning (RL) refers to the process whereby a learning system learns an associative mapping by maximizing a scalar evaluation (a reinforcement) of its performance from the environment. Delayed RL is a process in which the environment yields only a single scalar reinforcement collectively. Such tasks arise in the optimal control of dynamic systems and planning problems of artificial intelligence. Here, the authors `provide a comprehensive tutorial survey of various ideas and methods of delayed RL'. The connexion with stochastic optimal control is explored, and differences between delayed RL and dynamic programming methods are discussed. Model-based and model-free methods are examined, and general issues relating to the practical implementation of RL algorithms are noted.