Algorithms for Reinforcement Learning

Livro digital

Título:
Algorithms for Reinforcement Learning

Autor:
Csaba Szepesvári

Categoria:
Tecnologia > IA

Doador:
Raffaello D. N.

Sinopse:
Reinforcement learning turns decision-making into a learning problem, and this lecture draft opens with the core contrast between supervised feedback and partial, delayed rewards before moving straight into the structure of Markov decision processes. The table of contents shows that the book is organized around a clear progression from overview to MDPs, value functions, and dynamic programming, so the reader gets the full chain from problem setup to solution methods. From there, the material widens into the main families of RL algorithms: temporal-difference learning, every-visit Monte Carlo, TD(lambda), function approximation, gradient temporal-difference methods, least-squares methods, and the choice of function space. The control half then shifts to bandits, online and active learning in MDPs, Q-learning in finite and approximate settings, and actor-critic methods, which makes the scope broader than a simple introduction. What stands out is the book's unusually theoretical shape. The appendix on discounted Markov decision processes and Banach's fixed-point theorem signals a mathematically grounded treatment, not a lightweight survey. Readers looking for a compact but serious map of reinforcement learning will find both the algorithmic catalog and the limits of each approach.