Bandit Convex Optimisation

Tor Lattimore is a researcher at Google DeepMind working on reinforcement learning, bandits, optimisation and the theory of machine learning. He is the co-author of an introductory book on bandit algorithms and has published nearly 100 conference and journal articles. He is an action editor for the Journal of Machine Learning Research.

• Preface; 1. Introduction and problem statement; 2. Overview of methods and history; 3. Mathematical tools; 4. Bisection in one dimension; 5. Online gradient descent; 6. Self-concordant regularisation; 7. Linear and quadratic bandits; 8. Exponential weights; 9. Cutting plane methods; 10. Online Newton step; 11. Online Newton step for adversarial losses; 12. Gaussian optimistic smoothing; 13. Submodular minimisation; 14. Outlook; Appendix A. Miscellaneous; Appendix B. Concentration; Appendix C. Notation; Bibliography; Index.

'A landmark text on bandit convex optimization by an authority in the field. This book develops the full theory of zeroth-order online convex optimization-where one must learn from noisy function values without gradients-establishing regret bounds and presenting elegant algorithms from gradient descent to cutting planes, multiplicative updates, and Newton methods. Touching on all areas central to advanced optimization, it is an essential companion for researchers, offering both the conceptual foundations and the algorithmic toolkit that continue to drive progress in online convex optimization and mathematical optimization more broadly.' Elad Hazan, Princeton University