DeepRead
Cover of Reinforcement Learning and Optimal Control

Reinforcement Learning and Optimal Control

Dimitri Bertsekas · 2019

Machine learning, statistics and probability › Applied machine learning and computer vision

#13,556 most recommended

About the book

A graduate-level textbook on sequential decision-making under uncertainty, covering dynamic programming for deterministic and stochastic problems, finite and infinite horizon formulations, and the Bellman equation as their common foundation. It treats reinforcement learning as approximate dynamic programming and connects it to optimal control and model predictive control. Its central thesis is that many reinforcement learning methods can be unified as approximation in value space: a cost function is approximated offline, often by training a neural network or other architecture on simulated data, and then used in online lookahead minimization to select controls. It argues that rollout and policy iteration can improve on the approximation, giving near-optimal performance with bounded error. The book also presents this framework as the basis for successes such as TD-Gammon and AlphaZero-style game play, and as a bridge between control theory and modern learning-based methods.

Written by Claude from what it knows of this book.

What Hacker News says 50% recommend · 0% critical

00.512020202120222020 peak · 1
This bookAverage book in Applied machine learning and computer vision

In the commenters’ own words

What Open Library readers say

No Open Library reader has rated it yet.

Open Library’s page for the book

What critics say, via Wikipedia

No Wikipedia article about this book yet, so no critics to quote.

Reddit communities — coming later

Details

First published
2019
ISBN
9781886529397
Editions known
2 at Open Library
Open Library
OL29241210W

Artificial intelligenceDynamic programmingMathematical optimizationNeural networks (Computer science)Reinforcement learningScience

Explore in the mosaic

Read next: its neighbours on the mosaic