Practical Internship: Reinforcement Learning in Logistics

Boskalis

Papendrecht · Posted Nov 10

Apply

At a glance

Location: Papendrecht, Netherlands
Work type: Onsite
Category: Operations

Practical internship focusing on extending a reinforcement learning model for dike construction project optimization, using a DES+RL tool to improve KPIs like cost and emissions, with mentorship from software developers and AI specialists and potential continuation into an academic thesis.

What they are looking for

Reinforcement learningPythonPpoActor-criticMlflow

Location: Papendrecht, ZH, nl Department: Corporate development Employment Type: {'id': 'intern', 'label': 'Intern'}

Company Description

Working at Boskalis is about creating new horizons and sustainable solutions. In a world where population growth, an increase of global trade, demand for (new) energy and climate change are driving forces, we challenge you to make your mark in finding innovative and relevant solutions for complex infrastructural and marine projects.

Job Description

Are you looking for a practical internship where your work on reinforcement learning can contribute directly to real-life construction projects?

Many Boskalis projects involve reinforcing the dikes that protect the Netherlands against flooding and sea-level rise. These projects require large quantities of soil to be excavated, transported and reused. Planning these soil flows efficiently can reduce both project costs and emissions. We have developed an in-house soil-flow optimization tool for dike construction projects. Soil-flow planning is a complex problem involving different soil classes, quantities, locations and timing constraints. Engineers need to determine when and where excavated soils should be reused, stored, transported, sold or disposed of while meeting the project's construction requirements and optimizing objectives such as cost, transport, emissions and circular use of materials.

The tool combines a discrete-event simulation (DES) of the construction process with reinforcement learning (RL). The simulation represents the project environment, including supply and demand locations, timing, soil characteristics and logistics. Your challenge is to extend this existing model to a new dike construction project and train it to optimize new objectives, with a particular focus on KPIs like project cost or emissions. You will work with an existing model and codebase, real project data, and engineers to make the model perform on a real-world use case.

We also welcome a student who critically evaluates the existing modelling approach and identifies opportunities for improvement, for example through reward shaping, robustness metrics or hyperparameter tuning. If you demonstrate aptitude during this applied development assignment, your findings could provide the starting point for a follow-on academic thesis project, if you wish to continue the research.

Your main tasks will be

Understand the existing DES and RL architecture and how project constraints are translated.

Help configure and extend the existing model for a new dike project and environment.

Design and implement reward functions for cost, emissions or other relevant project KPIs, while ensuring that fundamental objectives such as completing the dike on time with available materials remain satisfied.

Train and evaluate Maskable PPO agent(s) and analyze the effect of reward shaping, hyperparameters and different optimization objectives.

Test and validate model performance against defined project use cases and benchmark scenarios, together with engineers and data scientists.

Use metrics and experiment tracking to understand why an agent learns a particular behavior rather than only reporting its final reward.

Document experiments and recommendations so that results can be reproduced and used in further development.

Your qualities and experience

You are currently studying at university level in Artificial Intelligence, Computer Science, Operations Research, Applied Mathematics or a closely related field.

You have practical or academic experience with reinforcement learning.

You understand the fundamentals behind PPO and actor-critic methods, rather than treating an RL algorithm purely as a black box.

You understand concepts such as reward functions, discount factors, exploration, training stability and hyperparameters.

You are comfortable programming in Python.

Experience with MLflow and Databricks is an advantage.

You enjoy debugging complex systems and interpreting why an ML model behaves as it does.

You can work independently but are comfortable discussing assumptions and results with engineers and data scientists.

Most importantly, you are critical and curious: when an experiment produces an unexpected outcome, you want to understand why and propose what to try next.

You do not need to be an expert in dike construction or soil logistics. We are looking for someone with strong RL foundations and the motivation to learn the engineering domain.

The following competences are also important

A good command of written and spoken English, currently living in the Netherlands.

Strong communication and collaboration skills as well as the ability to work independently.

A mindset geared towards continuous learning and curiosity for tackling new challenges.

Ability to handle feedback and reset your course if needed.

If you are excited about the opportunity to work on innovative projects and have the drive to make a significant impact, we would like to hear from you. Explain to us why your background and expertise make you the ideal candidate for our AI program and join us in shaping the future of Boskalis. Apply now and help us shape the future of our industry.