IE437
IE437 · KAIST Industrial & Systems Engineering

Data-DrivenDecision Makingand Control

One question — how do we decide well under uncertainty? — asked along three axes, answered by two lineages.

Follow a decision from a known model to learning from experience and from a fixed log. Each chapter includes a core learning route, a worked example, and a short problem to try. These are the interactive lecture notes — fourteen chapters you can open in a browser, step through, and run the algorithms inside.

three axestwo lineages14 chapters
Lecture materials → Midterm
drag, or focus and use the arrow keys, to rotate
01 · The course

One question, three axes, two lineages

Industrial decisions do not arrive as solved problems. They arrive as a messy situation, a budget, and a deadline — and turning that into something you can actually compute is the work. This course builds that ability from the ground up: state a decision problem, admit what you do not know about it, and then learn the missing part from data.

01

One question

How do we make good decisions under uncertainty? That is the entire course. Every lecture examines it under explicit assumptions: known models, uncertain quantities, sequential actions, and training from a fixed log. A shared temperature-control example connects the calculations.

02

Three axes

Any decision problem sits somewhere in a cube: is it one decision or a sequence; is the world given to you as a model or must it be learned from data; and do you decide alone or against others. The course is a route through that cube, and each lecture is one step along it.

03

Two lineages

Dynamic programming and optimal control provide complementary views of sequential decisions. Both use values, policies and Bellman reasoning. The course uses this comparison to connect known-model calculations with value-based, policy-based and model-based learning.

02 · Notices

Announcements

pinned

Midterm project is out — due Mon 2 Nov, 10:00

The midterm (30% of the grade) is a two-week individual take-home project, in the form of Assignment 1: a problem definition, the code behind it, an interactive demonstration and a proposal. A nationwide convenience-store chain must decide how many distribution centres to run and where. You formulate the daily routing of one centre as the lower level, connect it to the upper-level decision on the number and location of centres, search under an evaluation budget, and say explicitly how you balanced modelling the real operation against being able to compute an answer. The tools are those of Lectures 1, 2, 4 and 5.

Download the package in the Midterm section — read the brief's At a glance box and Section 5 (How to approach it) first. Start with a toy instance and have the whole pipeline running by day 7.

Released Mon 19 Oct 10:00 · Due Mon 2 Nov 10:00 KST · one ZIP (≤ 200 MB) per student, submitted on KLMS. Questions in class and by email to the TA; any answer that changes the brief will be posted here as a new version.

pinned

Lecture notes are online

All fourteen chapters are published on this page — open one in the browser to step through it interactively, or download the PDF of the same slides. Each chapter is a single self-contained file, so a downloaded deck keeps working with no network.

Announcements for the term will appear here.

03 · Midterm

The midterm project

30% of the grade · released Mon 19 Oct 2026, 10:00 · due Mon 2 Nov 2026, 10:00 KST · individual.

The midterm is a two-week take-home project in the form of Assignment 1: a problem definition, the code behind it, an interactive demonstration and a proposal. A nationwide convenience-store chain must decide how many distribution centres to run and where. You formulate the daily routing of one centre as the lower level, connect it to the upper-level decision on the number and location of centres, search under an evaluation budget, and say explicitly how you balanced modelling the real operation against being able to compute an answer. The tools are those of Lectures 1, 2, 4 and 5.

Read the brief's At a glance box and Section 5, How to approach it, first. Start with a toy instance and have the whole pipeline running by day 7. Submit one ZIP (at most 200 MB) per student on KLMS. Questions in class and by email to the TA; any answer that changes the brief is posted under Notices as a new version.

  1. Midterm Project

    Distribution-Centre Network Redesign

    Midterm project — formulating and solving a hierarchical decision

    Download ZIP

    Redesign a nationwide convenience-store chain's distribution-centre network. Formulate one DC's daily operation as the lower level, connect it to the upper-level decision on how many centres and where, search under an evaluation budget, and justify the balance between modelling the real operation and computing an answer — with the tools of Lectures 1, 2, 4 and 5. Two-week take-home, individual.

    Version
    2.2
    Due
    Mon, 2 Nov 2026 · 10:00 KST (UTC+9)
    Package
    4 Markdown files
04 · Assignments

Put the ideas to work

Download the complete assignment package, or choose an individual document. Start with the assignment brief for requirements and grading.

  1. Assignment 01

    Movie Programming

    Problem Definition and Decision-Support Design

    Download ZIP

    Define a movie-programming problem, build an interactive demo, and propose a decision-support approach for a cinema operator.

    Version
    2.1
    Due
    Fri, 9 Oct 2026 · 23:59 KST (UTC+9)
    Package
    4 Markdown files

    Worked examples · MD + interactive demo + proposal

    Read the completed Markdown definitions, present the English 44-slide and 59-slide proposals in their original designs, and experiment with the independent Response Model Simulator and latest movie-first M0 audience app. Each case exposes one English demo site and guide.

    Open worked examples Download examples ZIP
05 · Syllabus

How the term runs

2025 Fall
An upper-undergraduate course on data-driven decision making, with research extensions. Everything you need is on this page — the fourteen chapters below are the material, and there is no textbook to buy.
Instructor
Jinkyoo Park
E2-1, #4212
jinkyoo.park@kaist.ac.kr
Mon 16:00 – 17:00, straight after class · other times by appointment, by email
Teaching assistant
Inhyuck Song
son9ih@kaist.ac.kr
Term
2025 Fall
Lectures
Mon & Wed · 14:30 – 15:45
Location
(E2-2) Industrial Engineering & Management Bldg. #1501
Expected preparation
Calculus, linear algebra, basic probability, and Python
Textbook
None — these lecture notes are the material
Language
English

The course opens with optimisation and the fundamentals of Bayesian statistics, then spends its static half on making a design decision from data alone — Bayesian optimisation when you may still query the world, and surrogate-based and generative design optimisation when you may not. The dynamic half turns to decisions that unfold over time: Markov decision processes and dynamic programming, then value-based, policy-based, model-based and finally offline reinforcement learning.

The algorithms do not stay on paper. You will implement them, deploy them as APIs, and integrate them as AI agents through MCP servers — which is what the final project asks you to demonstrate.

Assessment

weights add to 100
  1. 40%

    Final project

    One industrial problem, carried from business context to a deployed prototype. The course's largest piece of work and the only one that leaves the classroom — it is set out in full below.

  2. 30%

    Midterm examination

    A two-week take-home project in the form of Assignment 1, over the first half of the course — optimisation modelling, Bayesian statistics, Bayesian optimisation, and design optimisation from a fixed dataset — applied to one hierarchical decision problem. A problem definition, an interactive demonstration and a proposal, with the methods prescribed. The brief and the dates are in the Midterm section.

  3. 20%

    Assignmentsdetails to come

    Two assignments, each run as a hackathon: a problem taken from a well-posed statement through to a working algorithm in a single sitting. They are rehearsals for the project's middle stage — formulate, solve, verify — at a scale you can finish in a day. Details are announced with each.

  4. 10%

    Attendance

    Lectures meet on Monday and Wednesday at 14:30 in the same room, and attendance is recorded.

40% · The final project

One industrial problem, carried the whole way

Industrial problems do not arrive as objective functions, and a verified answer that never reaches anyone has changed nothing. The project asks you to carry a single real problem across the whole distance — find it, formulate it, solve it with the methods of this course, and wrap the solution in something a person could actually operate. The scope is a working application prototype, not a notebook and not a report.
  1. 01

    Business context

    “Is this worth solving?”

    Begin where the problem lives rather than where the mathematics is convenient. Who decides, how often, and under what constraint? What does a better decision earn, and what does current practice already achieve? You finish this stage with a problem you can defend as worth solving and a baseline you will have to beat.

  2. 02

    Algorithm

    “Is the answer right?”

    Translate the situation into something computable — the objective, the levers, the constraints, and the part you do not know — then build the decision algorithm using an appropriate core method from this course. Compare it with the baseline under the same conditions, test against held-out evidence, and explain where it improves or fails. The project does not require implementing every research paper.

  3. 03

    Application

    “Does it reach the user?”

    Wrap the verified solution so somebody else can run it — an interface, an API, or an MCP server an AI agent can call — and deploy it where it can be reached. The prototype has to survive contact with its own operating loop: real inputs, an answer returned, and a next decision that follows from it.

What you hand in

  • A problem document

    The formulation and the case for it, kept up to date as the problem changes under you — not written the night before.

  • A deployed prototype

    Running where someone other than your team can reach it, with the algorithm behind it and a demonstration of the loop it closes.

  • A decision log

    What you asked the AI, what came back, and what you adopted or rejected — with the reason. Unverified AI output is not a deliverable.

Teams, milestone dates and the final showcase are announced during the term.
06 · Materials

Lecture notes

14 chapters · 1011 slides
Each chapter is a complete interactive deck and a downloadable PDF of the same slides. Start with the learning route, work through the main examples, and try the questions before revealing their answers. Research readings and detailed derivations follow in appendices.
MapThe whole course on one cube
00
Data-Driven Decision Making and Control
Deciding well under uncertainty — three criteria, four categories, four real systems
The whole course on one picture: a cube whose three axes are the three ways a decision problem gets harder. Every later lecture is placed at a corner of it, and four working systems — wind farms, traffic signals, a semiconductor furnace, a robot fleet — show what the methods decided in the field.
71 slides openPDF
Part IThe given world — one decision, a known objective
01
Optimization Problem Modeling
Before you can learn a decision, you must be able to state one
How to state a decision problem so that it can be solved at all: variables, objective, constraints, and the standard form everything later is written in. Convexity is the watershed — on one side a local optimum is global, on the other it is not — and the KKT conditions are how you certify an answer rather than merely find one.
93 slides openPDF
Part IIThe uncertain world — the objective becomes unknown
02
Fundamentals on Bayesian Statistics
From coin tossing to Bayesian regression
Follow the original lecture from coin tossing through conjugate models to regression. Work each model, try an experiment, and explain the result before moving on.
96 slides openPDF
03
Bayesian Networks
A belief about many things is a graph
A belief about many variables at once, drawn as a graph. Factorization is what makes an intractable joint tractable, a missing edge is a conditional independence you can read straight off the graph, and adding decision and utility nodes turns it into an influence diagram — a decision model that leads toward MDPs, and the seed of the entire second half of the course.
114 slides openPDF
04
Bayesian Optimization
Put the belief to work — act to learn, and to win
Now the belief acts. A Gaussian process carries uncertainty about an expensive unknown function, and an acquisition function turns that uncertainty into the next experiment worth running. This is the first policy in the course, and the exploration–exploitation dilemma it raises comes back in every chapter after it.
100 slides openPDF
Part IIIDesign from a fixed dataset — the oracle is gone
05
Data-Driven Design Optimization — Surrogate-Based
Deciding about designs no one has measured
The oracle is gone: a fixed dataset, and no way to test a new design. Fit a surrogate, optimise it, and the optimiser walks into the region where the surrogate is confidently wrong. Three research streams answer that failure — conservatism, uncertainty, and knowledge of the system.
75 slides openPDF
06
Data-Driven Design Optimization — Generative-Based
Learn where good designs live — then search, refit, condition or steer
A surrogate says how good a design is; it does not say what a design looks like or where to look. Learn a generative model of plausible designs, then let a performance signal in at one of four places — the coordinates the search moves in, the distribution designs are drawn from, the performance the generator is asked for, or each step of its sampler — and see what each assumes, solves and leaves.
78 slides openPDF
Part IVThe world that unfolds — two lineages, each losing its model
07
Markov Decision Processes & Dynamic Programming
A good plan stays good from wherever you land
The deepest crossing in the course: from choosing once to choosing at every state your own choices create. The Bellman equation says what a good plan is, and value and policy iteration solve it exactly — so long as somebody hands you the model.
55 slides openPDF
08
Value-Based Reinforcement Learning
Learning to act when the model is gone
Delete the model from Lecture 7 and the Bellman equation can no longer be solved, only sampled. Monte Carlo and temporal-difference learning, Q-learning, the cost of exploring, and the deadly triad that explains why adding function approximation makes the whole thing fragile.
56 slides openPDF
09
Optimal Control & Planning
The control-theoretic origin of sequential decision making
Sequential decision making's other parent. Control theory arrives where dynamic programming did by a different road — the HJB equation, LQR and its Riccati equation, Pontryagin's maximum principle — and hands back a feedback law instead of a table of values.
55 slides openPDF
10
Policy-Based Reinforcement Learning
Optimal control with the dynamics replaced by data
Lecture 9 with the dynamics replaced by data. Optimise the policy directly by gradient ascent, cut the variance with a baseline, and control the size of policy updates — the machinery that PPO is built from.
61 slides openPDF
11
Model-Based Reinforcement Learning
Put the model back — learned — and let the two lineages reunite
Put the model back, but learn it. A learned model can be planned with, imagined through, or differentiated for design — and wherever it is wrong, rollouts drift away from reality. This is where the two lineages of the course rejoin.
67 slides openPDF
FinaleInteraction itself is taken away
12
Offline Reinforcement Learning
You cannot try. Learn from what was already done.
The last thing taken away is the right to act. Learn a policy from a fixed log with no exploration at all, where any action outside the data is a guess the Q-function cannot check — the same failure Lecture 5 met, answered the same way, with conservatism.
50 slides openPDF
AppendixThe toolbox underneath
A
Probability Review
The toolbox beneath the course — and where each tool was used
The toolbox the course leans on: expectation and its linearity, variance and the bias–variance split, Bayes' rule, and the Gaussian. Each one is shown at the point where a lecture actually leaned on it.
40 slides openPDF
Open presents the chapter in the browser — it is interactive, with click-through reveals and live simulators. PDF downloads the same chapter as slides. In a deck: → next reveal · ← back · ↑↓ skip a slide · M index · P print · F fullscreen · ? all keys. Each chapter is a single self-contained file: save it and it runs offline.