Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Syllabus · Plan de cours

IFT 6162: Apprentissage par renforcement, commande optimale (Reinforcement Learning, Optimal Control)

English
Français

Course information

Course codeIFT 6162
TitleReinforcement Learning, Optimal Control (official registrar title is in French: Apprentissage par renforcement, commande optimale)
CreditsTBD
TermFall 2026
ScheduleMondays 10:30–12:29 and Thursdays 13:30–15:29; Aug 31 – Oct 16 and Oct 26 – Dec 9, 2026 (14 teaching weeks, ~25 sessions)
RoomCampus Montréal, room TBD
InstructorPierre-Luc Bacon
EmailTBD
Office hoursTBD
Teaching assistantsTBD
Language of instructionTBD (course materials are in English)
Course websitehttps://pierrelucbacon.com/rlbook/

Description

This course builds reinforcement learning up from its foundations in dynamics, optimization, and control. Rather than starting from the tabular Markov decision process and working outward, we start from the decision problem itself: how to write down a model of a system, how to state an objective that unfolds over time, and how to solve the resulting optimization problem numerically. Dynamic programming, model predictive control, and modern deep RL algorithms then appear as different answers to the same question, each carrying its own assumptions and failure modes.

The organizing theme is that most of the work in applying RL happens before an algorithm is chosen. Sensors produce noisy data, constraints are non-negotiable, and objectives shift or conflict. A practitioner who can only reach for a policy gradient method is poorly equipped for that reality. A practitioner who recognizes the same mathematical structure across trajectory optimization, MPC, dynamic programming, and deep RL can choose the right tool and explain why.

Reinforcement learning did not develop in isolation: its foundations draw on control theory, dynamic programming, operations research, and economics. The same ideas recur under different names in each community. Making those connections explicit is a central goal of the course.

Prerequisites

Students are expected to be comfortable with:

  • Mathematics. Linear algebra, multivariable calculus, and probability at the undergraduate level.

  • Optimization. Unconstrained and constrained optimization; gradient-based methods. Familiarity with KKT conditions is helpful but will be reviewed.

  • Programming. Python, with working knowledge of NumPy. The course uses JAX and nonlinear programming solvers; no prior experience with either is assumed.

  • Machine learning. A prior course covering supervised learning and neural network training.

Prior exposure to reinforcement learning is not required.

Learning outcomes

By the end of the course, you should be able to:

  1. Formulate a sequential decision problem as a discrete-time optimal control problem or an MDP, stating dynamics, objective, and constraints explicitly, and justify the modeling choices you made.

  2. Solve trajectory optimization problems numerically using single shooting, multiple shooting, and direct collocation, and explain the trade-offs among them in terms of conditioning, sparsity, and solver behavior.

  3. Implement model predictive control, and reason about recursive feasibility, stability, constraint softening, and fallback behavior.

  4. Derive dynamic programming recursions for finite- and infinite-horizon problems, and characterize Bellman operators as contractions on an appropriate space.

  5. Explain approximate dynamic programming as the projection of a Bellman residual onto an approximation space, and connect this view to fitted value and fitted Q iteration.

  6. Analyze the Monte Carlo estimators used throughout RL, identifying sources of bias and variance, including maximization bias.

  7. Situate deep RL algorithms, including DQN and its extensions, DDPG, TD3, path consistency learning, MPPI, and policy gradient methods, within this framework, and explain each design choice as a response to a specific difficulty.

  8. Read and critically assess research papers spanning the RL, control, and operations research literatures.

Textbook

The primary text is the course book, written for this class and freely available online:

Pierre-Luc Bacon. Building Up RL: From Dynamics and Control to Learning. https://pierrelucbacon.com/rlbook/

The book is executable: figures and examples are generated from code you can run and modify. The source is on GitHub; corrections and issues are welcome.

Supplementary references

These are recommended for depth on particular topics; none is required.

  • D. P. Bertsekas. Dynamic Programming and Optimal Control, Vols. I–II. Athena Scientific.

  • M. L. Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley, 1994.

  • R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction, 2nd ed. MIT Press, 2018.

  • J. T. Betts. Practical Methods for Optimal Control and Estimation Using Nonlinear Programming, 2nd ed. SIAM.

  • J. B. Rawlings, D. Q. Mayne, and M. M. Diehl. Model Predictive Control: Theory, Computation, and Design, 2nd ed. Nob Hill.

  • J. Nocedal and S. J. Wright. Numerical Optimization, 2nd ed. Springer.

  • W. B. Powell. Reinforcement Learning and Stochastic Optimization. Wiley, 2022.

Evaluation

ComponentWeightDue
Applied project30%TBD
In-class midterm 115%TBD
In-class midterm 215%TBD
In-class final examination40%TBD
Total100%

The three examinations are completed in person on paper, without electronic devices or generative-AI tools. Each student may bring one double-sided reference sheet that they prepared themselves. Questions may ask students to derive or interpret a result, inspect supplied code, identify a flaw in a model or algorithm, and justify a diagnosis. Memorizing Python syntax is not an examination objective.

Exercises, self-checks, and coding practice are formative and ungraded. They prepare students for the analytical and diagnostic work required on the examinations and in the project.

The applied project is completed in teams of exactly three. If enrollment makes that impossible, exceptions will be arranged privately by the instructor. Each team will formulate and investigate a substantive control-and-learning problem rather than reproduce a canned benchmark. The deliverables are working code, a poster, and an oral defense. A short proposal is an ungraded checkpoint. The team artifact establishes the base project grade; an individual’s grade may be adjusted when the oral defense shows a materially different level of understanding.

Grading scale. TBD

Project checkpoint and late policy. Dates and late arrangements are TBD.

Tentative schedule

The course meets twice a week, Mondays 10:30–12:29 and Thursdays 13:30–15:29, over 14 teaching weeks, for roughly 25 sessions in total. Each row below therefore covers two meetings, except where a holiday intervenes. Pacing is indicative and will be adjusted as the term goes on.

WeekTopicReading
1 · Aug 31 – Sep 3Course overview; system boundaries; state and actionControlled Systems
2 · Sep 7 – 10Dynamics models and the state-space perspectiveControlled Systems
3 · Sep 14 – 17Stochastic dynamics; partial observation; programs and data as model interfacesStochastic Dynamics, Model Interfaces
4 · Sep 21 – 24Discrete-time optimal control problems; existence and optimality conditionsFinite-Horizon Optimal Control
5 · Sep 28 – Oct 1Adjoints and the Pontryagin principle; single and multiple shootingDiscrete-Time PMP, Numerical Trajectory Optimization
6 · Oct 5 – 8Direct transcription; polynomial interpolationContinuous-Time Collocation
7 · Oct 12 – 15A compendium of direct transcription methods; worked examplesContinuous-Time Collocation
8 · Oct 26 – 29Closing the loop by replanning; theoretical guarantees; MPC variantsReceding-Horizon Control, MPC Variants
9 · Nov 2 – 5MPC failure handling; parametric optimization and approximate controllersReliable MPC, Parametric Controllers
10 · Nov 9 – 12Backward recursion, continuous spaces, and the linear quadratic regulatorFinite-Horizon Dynamic Programming
11 · Nov 16 – 19Stochastic and infinite-horizon MDPs; Bellman operators; value and policy iterationStochastic DP, Infinite-Horizon MDPs
12 · Nov 23 – 26Regularized MDPs; weighted residuals; approximate Bellman equationsRegularized DP, Weighted Residuals, Approximate Bellman Equations
13 · Nov 30 – Dec 3Monte Carlo methods and overestimation bias; fitted Q iteration, NFQ, DQNMonte Carlo, FQI
14 · Dec 7Amortized action optimization; stochastic gradient estimators; regularized and direct policy optimizationAmortized Action Optimization, Gradient Estimation, Regularized Policy Learning, Policy Gradients

No class on:

  • Monday September 7: Labour Day (Fête du travail)

  • Monday October 12: Thanksgiving (Action de grâce)

  • October 19–25: période d’activités libres (no meetings; this is the gap between the two blocks in the registrar’s listing)

  • Thursday December 10 falls outside the course’s end date of December 9, so week 14 has a Monday session only.

The appendices on worked examples, initial value problems, and nonlinear programming are reference material used throughout the term rather than assigned to a specific week.

Course policies

Attendance. TBD

Collaboration. Discussion and collaboration are encouraged on formative work. The project is collaborative within the assigned team. Each student must be able to explain and defend the team’s complete submission.

Generative AI. Generative AI may be used for formative work and the applied project. Disclosure is not required. Each student remains responsible for every claim, equation, baseline, experiment, and line of submitted code and must be able to defend them during the oral defense. Generative AI and electronic devices are prohibited during in-class examinations.

Academic integrity. All work is subject to the University’s regulations on plagiarism and fraud. See https://integrite.umontreal.ca/ for the full policy.

Accommodations. Students registered with the Soutien aux personnes étudiantes en situation de handicap (SESH) service should contact the instructor early in the term so that arrangements can be made. See https://vieetudiante.umontreal.ca/a-propos/service/soutien-personnes-etudiantes-situation-handicap or write to soutienhandicap@sve.umontreal.ca.

Changes to this syllabus. This document may be revised during the term. Changes will be announced in class and reflected on this page.