Research questions
- How can the three magic lifts best be modelled?
- What data is necessary to identify the unknown parameters of the system model?
- How can the system model be validated?
- What are the system model’s uncertainties, and how can they be quantified or bounded?
- How can the identified model be used in optimization, and is there a strategy that will ensure fairness among players?
Methodology
This thesis will take you on a journey from the ground floor of data acquisition about the lifts’ positions and demands. Then you will be lifted through the concepts of stochastic processes to model the arrivals, Markov Decision Processes and Finite State Machines for identifying and modelling the two collaborative lifts, arriving at queuing theory to compute the waiting times. Finally, the door is open for stochastic optimization to answer the key questions.
These classical approaches should be compared with taking the staircase of reinforcement learning and stepwise improve a deep learning model of the lifts' behaviours in response to the stochastic user demands. An issue with a deep learning model is the lack of quantified confidence in its results, and it can be difficult to assess the uncertainties. It could therefore be better to develop a multi-agent model of the lifts and the behaviour of the people at IFI, and then use ensemble simulations to address the uncertainty in your optimized solutions.
What you will learn
For the first part of the journey you will need some prior knowledge about data collecting devices, like a camera, how to program these, and possibly image classification by machine learning algorithms. You will learn:
- How to design and build a data acquisition system,
- How to store and handle potentially sensitive data,
- How to process and analyse the detected events.
For the last parts of the journey you will learn advanced scientific methods for
- Modelling and identification of stochastic systems over finite state spaces,
- Deep reinforcement learning for state representation learning, or
- Multi-Agent system modelling and Monte Carlo methods for system simulation, and
- Optimization of stochastic systems as multi-player games.
No prior knowledge is necessary in these areas, but you must appreciate some advanced mathematics. This journey is not for the faint-hearted!