timefold.ai @tomcools.be
LLMs can't optimize schedules...
but AI can.
by Tom Cools
GenAI Is Awesome 🚀
Solution for EVERYTHING 🤯
or is it? 🤔
What is something
LLMs can't do?
Can LLMs solve planning problems?
7 days
5 shifts per employee
30 employees
150 shift assignments
5 constraints (skills, time between shifts, etc.)
How do LLMs solve
my planning problem?
LLMs 🚶♂️🌳 ☀️
Can reasoning models optimize a realistic schedule?
That is "slightly" over-optimistic
Why do LLMs fail here?
Scheduling is not a sequence of decisions
30 employees
150 shift assignments
How many different solutions?
1 shift: 30 options
2 shifts: 30*30 options
3 shifts: 30^3 options
150 shifts: 30^150 options
150 shifts: 30^150
3.69 × 10^221
Atoms in the observable universe: 10^80
10^221: 141 more zeroes
You can't optimize for such a search space
by predicting the next token
nor by reasoning like a human.
LLMs can't optimize schedules
... but AI can!
Machine Learning
Machine learning is a branch of AI where systems learn patterns from data to make predictions or decisions without being explicitly programmed.
Machine Learning is Awesome 🚀
Can Machine Learning be used
to optimize schedules?
Learn from data...
but what data?
Mathematical Optimization
Group of techniques to find the ideal inputs from a set
in order to maximize or minimize a real function.
$\begin{aligned}
& \quad \sum_{s \in S} \sum_{e \in E} \mathrm{pref}(s,e) \, a_{s,e} \\
\end{aligned}
$
PlanningAI
Planning optimization made easy
- Optimization algorithms in domain code
- Open Source (Apache license)
- Java / Kotlin / Python
Demo
🖊 Model domain
⚖️ Define constraints
🚀 Run!
Example Usecases
Does Optimization Matter?
Vehicle Routing Case
Expected: -1% driving time
Result: -25% driving time
⇒ -10 million kg CO² emission per year
⇒ -100 million $ cost per year
LLMs can't optimize schedules
but AI can!
Future of AI?
Let LLMs write
the code
Integrate with
Model Context Protocol (MCP)
| Technique | Strengths |
|---|---|
| GenAI | Data extraction |
| Machine Learning | Prediction |
| PlanningAI | Scheduling |
To Conclude
GenAI Is Awesome 🚀
...but it can't do everything...
Tom Cools: (@tomcools.be)
Feedback through VJUG Slack appreciated!