Machine learning and stochastic control

Cours Master 2 Probabilités et Finance, Sorbonne Université et Ecole Polytechnique 

Planning 2025/2026

Cours: 

Mardi 13 janvier 2026, 9h-12h,  salle 101, tour 15/25

Jeudi 15 janvier 2026, 9h-12h, salle 113, tour 16/26

Mardi 20  janvier 2026, 9h-12h,  salle 101, tour 15/25

Mardi 10  février 2026, 9h-12h,  salle 102, tour 15/25

Jeudi  12 février 2026, 9h-12h, salle 106, tour 14/15

TP: (Samy Mekkaoui et Alexandre Alouadi)

Mardi 27 janvier 2026, 9h-12h, salle 102, tour 15/25 

Mardi 3 février 2026, 9h-12h, salle 102, tour 15/25 

Jeudi 19 février 2026, 9h-12h, salle 106, tour 14/15 

This course will present some recent developments on the interplay between optimal control and machine learning with emphasis on continuous time setting and financial applications.   It will be illustrated with some practical works.  Documents, slides and Labs posted on Moodle page

Part I: Foundations 

1. Introduction to stochastic control and machine learning 

• Overview of stochastic control

• ML meets stochastic  control

• Applications: finance, energy systems, robotics

2. Basics of Reinforcement learning (RL)

• Fundamental concepts: MDP, value functions, Bellman equation

• Value-based and policy-based algorithms

• Transition to continuous-time RL: limitations of discrete time framework

3. Continuous time stochastic control RL

• Control with randomized policies

• HJB with entropy regularizer 

• Gibbs measure for the optimal randomized policy

Part II: Machine learning techniques for control 

4.  Neural networks algorithms for PDEs and HJB equations

• Deep Galerkin and physics-informed NN

• Deep BSDE

• Deep backward dynamic programming scheme

5. Policy gradient methods in continuous time

• Policy gradient representation

• Actor/critic algorithms

6.  Q-learning and approximations in continuous time

• q-learning and Hamiltonian function

• q-learning actor/critic algorithms

Part III: Generative modeling with Schrödinger bridge diffusion 

7. Fundamentals of Schrödinger bridge (SB)  and connection to optimal transport/control 

• Introduction to SB 

• Solution to SB, Schrödinger system and Sinkhorn algorithm

8. SB as generative model

• Deep generative learning 

• Optimal interpolation for time series 

References 

• Chen, Georgiou, Pavon, Stochastic control liaisons: Richard Sinkhorn meets Gaspard Monge on a Schrödinger bridge, 2021, SIAM Review, 63(2).

• De Bortoli, Thornton, Hen, Doucet: Diffusion Schrödinger bridge with applications to score-based generative modeling, 2021, NeurIPS

• Germain, Pham, Warin: Neural networks-based algorithms for stochastic control and PDEs, 2023, Machine learning and data sciences for financial markets: a guide to contemporary practices, CUP,  eds. A. Capponi and C-A. Lehalle.

• Hambly, Xu, Yang: Recent advances in Reinforcement learning in finance,  2021, arXiv:2112.04553

• Hamdouche, Henry-Labordère, Pham: Generative modeling for time series via Schrödinger bridge, 2023, arXiv:2304.05093

• Jia, Zhou: q-Learning in continuous time, 2023, JMLR

• Sutton, Barto: Introduction to reinforcement learning, 2nd edition, 2016. 

• Wang, Jiao, Xu, Wang, Yang: Deep generative learning via Schrödinger bridge, ICML, 2021.  

• Wang, Zariphopoulou, Zhou: Reinforcement learning in continuous time and space: a stochastic control approach, JMLR, 2020, 21, 1-34