• Nowe, Ann (Administrative Promotor)
  • Salvant, Thierry (Industrial Promotor)


The aim of this project is to control and guide the emergent behavior of a large scale learning multi-agent systems by incorporating expert knowledge. The use of reinforcement learning which lends itself very well for incorporating domain knowledge of various forms, seems a very natural way to control a large range of intelligent distributed systems, such as, air traffic control operations, smart energy grids, intelligent traffic signs, just to name a few.

The final goal of this system is to build a global plan aimed to solve a challenging multi-agent coordination problem. Several techniques to speed up and guide the learning for the single agent setting exist already, but most of these approaches have not been much explored yet in a multi-agent setting. Another focus will be on how to make the learning behavior of this setting be progressively relaxed, guaranteeing the viability of the system in all circumstances. Robustness and adaptability are also key factors, in order for the overall solution be stable against unexpected events.

In the end, the global plan will be communicated to a human user via a decision support system. The communication with the human controller is crucial because the responsibility stays with that person. Therefore it is important that the user is able to understand and interact with the proposed plan. This requires naturally an approach in which the emergent behavior of the multi-agent system can be engineered. This, in our case, will be explored by the generation and use of initial policies through an inverse reinforcement learning approach and reward shaping. The task to guide the multi-agent learning process and assist it in finding good policies in large scale multi-agent systems will be based on FCQ-learning, which will be extended to incorporate the required features.

The techniques developed within this project will be evaluated in the air traffic control and airport ground operations domain. They will be compared against the planning proposed by state-of-the-art decision support systems used in operational mode nowadays for the management of airport aircraft departure ground operations, referred to as Departure MANager (DMAN). To reach our evaluation goal, a prototype of a DMAN integrating the proposed approach will be developed and compared to a DMAN using the current state-of-the-art techniques. This comparison will be performed inside a testbed using fast-time simulation techniques.
Effective start/end date1/01/1331/01/17

    Flemish discipline codes

  • Applied mathematics in specific fields

    Research areas

  • Databases, Evolution of language, Programming languages, Mobile Computing, Artificial Intelligence, Serious games, Web systems, Software agents

ID: 3509659