The normal prerequisite for all projects below is that you are a student doing your master's in computer science at Delft including at least the course Advanced Algorithms (or Artificial Intelligence Techniques) and one other course in our group (e.g. the course Algorithms for Intelligent Decision Making gives a very good preparation for doing your final research project). If you are doing your master's at Delft in another subject, please additionally send us a motivation why you would like to make your thesis with our group.
Below you can find a representative sample of projects. The first table lists mostly applied projects (often with an external collaboration or internship), and the second table more fundamental projects and their source (typically industry or a scientific collaborator). If you are interested in one of these projects and meet the criteria above, please get in touch with the respective contact person in the Algorithmics group. If you meet the criteria but do not have a specific project in mind, please get in touch with the group secretary Kim Roos.
Applying AI Algorithms to Real-World Problems
Many exciting master's projects can be defined by applying state-of-the-art planning, scheduling, and coordination algorithms to real-world problems.
Power to Heat – Optimize use of industrial electric boiler
possible internship at Uniper
Traffic flow optimization
Delft Transport Institute
DEAL services, R. Negenborn (3mE)
AI-based traffic light control
possible internship at Siemens (Zoetermeer)
|Developing a Driving Simulator Environment for Studying Mixed Traffic||Transport & Planning (CEG)||Spaan|
Optimization of grid investments
possible internship at Alliander
Planning algorithms for energy grids
possible internship at Alliander
Dynamic, temporal planning on ships of the Royal Netherlands Navy
possible internship at TNO (The Hague)
possible internship at WithTheGrid
|Online/rolling horizon optimal bidding||possible internship at Jedlix, project with GreenChoice||de Weerdt|
|Charging electric car-sharing cars||possible internship at Jedlix||de Weerdt|
possible internship at Transvision
possible internship at NS/NedTrain
possible internship at NS/NedTrain
possible internship at NS/NedTrain
Effect and design of a market for flexibility in congested networks
possible internship at AgroEnergy
|Better house allocation for social housing||AMS, City of Amsterdam, UvA||de Weerdt|
Shift and personnel rostering with constraints
possible internship at KLM
|Applied optimisation with commercial data||possible internship at ORTEC||Yorke-Smith|
|Hybrid optimisation for electroplate assembly lines||possible internship at leading supplier||Yorke-Smith|
Comparing centralised and decentralised optimisation for dynamic pickup and delivery
|Planning for safety and security of critical industrial facilities||P. van Gelder (TPM)||Yorke-Smith|
Deep learning for combinatorial optimisation
Optimisation or machine learning in container port operations
possible internship at TBA Group
Engineer dispatching using hybrid machine learning
possible internship at Macomi
possible internship at The Missing MIddle
Agent-based modelling of housing market evolution
I. Nikolic (TPM)
Fundamental Challenges in AI Algorithms
A scientifically-important master's project can come from addressing a fundamental algorithmic challenge.
Multi-objective mechanism design
Multi-party multi-issue negotiation versus matching
C. Jonker (Interactive Intelligence, EWI)
Voting with incomplete preferences
University of Southampton
Auctions and tenders with uncertainty
|Learning to search||L. Bliek||de Weerdt|
|Long versus short-term decision making||G. Neustroev||de Weerdt|
|Decision diagrams for optimization||P. van den Bogaerdt||de Weerdt|
Robust local search using scenario addition
Experimental comparison of algorithms for (temporal) planning
J. Shah (MIT)
Data mining output of agent-based simulation models
I. Nikolic (TPM)
|Unit testing for functional agent-based languages||E. Chappin, I. Nikolic (TPM)||Yorke-Smith|
Influence of behavioural factors in maritime customs processes
F.J. Srour (Lebanese American University)
Optimization under culture
G.J. Hofstede (Wageningen)
Scheduling with disjunctions and preferences
B. Venable (Tulane)
|Approximate negotiation in international trade||C. Jonker (Interactive Intelligence, EWI)||Yorke-Smith|
Agent-based airport surface traffic planning under uncertainty
The amount of traffic at airports increases with every passing year. One of the major challenges which airports face is how to manage ground movements of aircraft in an effective and efficient way. Traditionally, to address the challenge of optimal airport surface traffic planning, approaches from the area of Operations Research have been used. However, such approaches have scalability issues with handling realistic amounts of traffic at large modern airports. In this project, agent-based planning approaches will be attempted to address the scalability issue. In contrast to traditional OR techniques, in multi-agent systems planning is performed by a large number of agents, which possess local information about the system and are able to communicate with each other.
The aim of the project is to develop an agent-based airport surface traffic planning algorithm in the context of a case study at a real airport (e.g., Schiphol). The algorithm should be able to deal with uncertainties, in particular in input parameters such as the expected landing time of aircraft and the expected pushback time of aircraft. The planning should ensure the scheduled runway arrival times for each departing aircraft.
This project is a collaboration with Aerospace Engineering. Contact: Matthijs Spaan
Fleet management at DEAL services
DEAL services has a platform for logistic services that mainly consists of information and communication software, e.g., planners, drivers and customers have an app with estimated arrival times, route information, etc. In the current approach, only the order of tasks for the individual vehicles and the assignment of recurrent tasks to vehicles are determined automatically; assignment of other tasks to vehicles is done manually, although efficient methods exist for the Vehicle Routing Problem (VRP). The assignment is to answer the question how efficient vehicle routing methods can be incorporated in the DEAL services platform?
In collaboration with DEAL services and R. Negenborn (3ME). Contact: Matthijs Spaan
Optimizing supply temperatures in district heating grids
A district heating grid is used to heat buildings and to provide them with hot tap water. This heat is produced centrally at one or more producers and then transported as hot water to consumers through insulated pipelines. Minimizing production costs while fulfilling the energy demand is a trade off between mass flow and supply temperature. A higher mass flow results in higher pumping costs, and higher supply temperature leads to higher energy losses during transport. The optimization has some challenging non-linearities, most notably the variable temperature propagation time.
This project is in collaboration (or possible internship) with Withthegrid. Withthegrid is a young and fast growing company that focuses on improving energy grids. It does this by designing and delivering internet of things monitoring devices and developing optimisation software. Current customers include Stedin, Eneco and Ennatuurlijk. Our office is currently based in an energy production facility giving a proper industrial look&feel. During your Master thesis/internship you will be working closely with the team and discuss on a daily basis with the engineering team members. Lunches are together and once a month there are Friday afternoon beers with the team.
Contact: Mathijs de Weerdt
Quality personal transport for elderly and disabled people
Currently the quality of personal transport for elderly and disabled people is very low. The main reason is the heavy competition for the three-year long contracts with governmental institutions. What if people can choose their preferred company per ride? What if people can choose between different pick-up times for their transport?
In this project you will test multiple methods for assigning transportation jobs to taxi companies, such that the preferences and quality for the end-users increases without increasing the costs for the government too much.
Contact: Mathijs de Weerdt
Improving maintenance of trains at NS
NS Maintenance (NedTrain), part of the national Dutch Railway group (NS group) is responsible for the tender, maintenance and revision of rolling stock. Trains to be maintained are delivered by NS Travelers, also part of the NS group.
Arrival of trains (release times of maintenance jobs) is quite unpredictable. NS Travelers decides exact supply of rolling stock at the very last moment and delays in the schedule occur on a daily basis. The exact duration of maintenance jobs is only known after initial inspection jobs, therefore there is significant uncertainty in the operational scheduling; Also, maintenance is performed by a number of teams at different locations, and each team constructs its own operational schedule. Furthermore, shunting at service sites, and crew assignment further complicate the problem to be solved.
Several master's thesis projects are possible with an internship at NS/NedTrain in Utrecht, on the following topics:
- Design a fast scheduling algorithm for the maintenance problem such that rescheduling is possible in real-time;
- Compute capacity of maintenance locations (nodes);
- Find a method to decompose the maintenance scheduling problem into a number of independent subproblems.
- Use machine learning methods to improve exact methods for solving the combined scheduling and shunting problem.
- Apply transfer learning to extend policies for smaller shunting yards to larger ones.
- Take preferences of maintenance engineers, global cost objectives and preferences of other parties into account when scheduling.
- To determine whether found schedules are any good, we would like to compare the costs to an optimal solution. Typically, however, optimal solutions cannot be found. However, we can have lower bounds (an optimal solution to a relaxed problem). Earlier work has shown how to obtain these for an abstract problem. Next steps are to use these in finding solutions, and to extend these to the realistic setting.
Planning for safety and security of critical industrial facilities
How do we design an industrial facility for safety and security? Or how do we retrofit an existing facility to mitigate safety risks and improve response to incidents? How do we ensure fail-safe/fail-secure for the most critical facilities? This thesis will study how to optimize an industrial facility layout for safety, using a combination of bayesian networks, constraint-based models and machine learning.
This project is a collaboration with TPM. Contact: Neil Yorke-Smith
Matching uncertain demand and supply in smart electricity networks
Integration of renewable energy in power systems is a potential source of uncertainty, because renewable generation is variable and may depend on changing and highly uncertain weather conditions. An example is power generated by wind turbines. Although renewable wind energy is clean and cheap, it may be intermittent and its availability is uncertain and difficult to predict. To reduce peak power consumption and to mitigate the effects of uncertain renewable power supply, electricity usage can be deferred in time such that demand and supply are balanced. An example is the charging process of an electric vehicle, which often does not have to be charged immediately, as long as the available power in the battery is sufficient to reach a destination. In our group we develop efficient planning techniques to automatically coordinate deferrable loads, such that electricity is used when renewable supply is available.
Various MSc projects can be formulated in this domain. Examples include, but are not limited to, coordination of charging electric vehicles, scheduling household appliances, electricity usage within network constraints and predicting electricity demand and supply. We also have an existing collaboration with the Electrical Sustainable Energy department of the EEMCS faculty.
Planning under uncertainty
Planning under uncertainty is a technique for enabling agents to successfully plan their decisions in domains with stochastic transitions and partial observability, e.g., because of noisy sensors. In the Artificial Intelligence community the Partially Observable Markov Decision Process (POMDP) and related multiagent models have become popular choices to address these hard planning problems. In the Algorithmics group a considerable body of expertise, algorithms and software is present regarding these methods. This allows for defining challenging projects focused on improving state-of-the-art algorithms with a high potential for publications.
Contact: Matthijs Spaan
Reinforcement learning for transportation, traffic and logistics
Reinforcement learning is a branch of machine learning focusing on agents that are able to automatically learn how they should act in their environment, and has been applied for coordination of multiagent systems in several real-world domains. For instance, reinforcement learning has been used for intelligent control of multiple traffic lights in the urban area, as well as optimization of traffic flow on highways. Other examples in the area of transportation and logistics include air traffic management and automated unloading of ships in a harbor. Typically, a real-world problem is modeled as reinforcement learning problem and then evaluated through realistic simulations.
MSc projects may focus on the application of reinforcement learning algorithms to realistic domains, but also allow for more fundamental research related to reinforcement learning and decision making under uncertainty. In our group expertise and software are present for reinforcement learning in transportation, traffic and logistics domains. However, students may also choose their own problem domain to define an interesting thesis project.
Developing a Driving Simulator Environment for Studying Mixed Traffic
The automotive industry jointly with many software companies and automotive suppliers are boosting their efforts to transform the future of transportation by making self-driving cars become a reality. Testing the interactions between automated and human driven vehicles in real life is complex as limited number of automated vehicles drive on our road network. Therefore, testing these interactions in a driving simulator environment in the first stage is more realistic. However, a suitable driving simulator environment which can simulate human driven vehicles and automated vehicles does not exist. Your aim in this thesis is to develop the driving simulator environment, including the integration of existing behavioural models for humans as well as automated vehicles from the literature into the driving simulator environment.
Acquiring CSP models with interactive machine learning
Constraint Programming is known as a powerful approach to combinatorial optimisation problems. Part of the "Holy Grail" of computing is for the user to state the problem and the computer to solve it. Progress over the last 30 years in CP means the computer can, for many problems, solve a CP model without human intervention. However, acquiring a suitable CP model still requires human expertise. This thesis project explores the recent uses of machine learning to acquire improved CP models.
Contact: Neil Yorke-Smith
Machine learning coupled with state-of-the-art OR algorithms
Reinforcement Learning (RL) has revolutionised what computers can do: game playing, speech translation, car driving, making artwork. Given their success, deep RL techniques are now being applied to combinatorial optimisation problems, such as vehicle routing . These kind of problems are important in business, science and policy making. Traditionally their study has been called Operations Research (OR) .
The aim of this project is understand how RL can be used as an alternative and as a complement to already successful OR models and methods. Relevant questions could for example be:
- What is the latest work on deep neural networks to solve OR problems?
- How can RL and OR/AI work in synergy?
- How can machine learning help to acquire models, especially constraint-based models?
- How can machine learning help to solve models, especially constraint-based models?
- What are the limitations of deep RL for combinatorial optimisation?
Contact: N. Yorke-Smith
 Wouter W. M. Kool, Max Welling: Attention Solves Your TSP. CoRR abs/1803.08475 (2018)
 Leo Liberti, Thierry Marchant, Silvano Martello: Twelve surveys in operations research. Annals of OR 240(1): 3-11 (2016)
 Michele Lombardi, Michela Milano, Andrea Bartolini: Empirical decision model learning. Artificial Intelligence 244: 343-367 (2017)
 Andrea Galassi, Michele Lombardi, Paola Mello, Michela Milano: Model Agnostic Solution of CSPs via Deep Learning: A Preliminary Study. CPAIOR 2018: 254-262
 Christian Bessiere, Frédéric Koriche, Nadjib Lazaar, Barry O'Sullivan: Constraint acquisition. Artificial Intelligence 244: 315-342 (2017)
Unit testing for functional agent-based languages
NetLogo is a functional multi-paradigm programming language for agent-based social simulation. NetLogo is widely used is the social sciences. It is open source, based on Scala. NetLogo has some features for static and dynamic code analysis and a simple profiler. It has no features for unit testing or advance code analysis. This project studies the requirements, feasibility and implementation of a prototype for NetLogo.
This project is a collaboration with TPM faculty. Contact: Neil Yorke-Smith
The topics listed above are only a sample of possible projects; fully up-to-date information can be obtained by contacting our staff members:
If you are currently not studying in Delft but would like to study here, you will find TU Delft a stimulating academic and social environment. To apply for a MSc place, please read and follow the formal guidelines of Delft University of Technology.