The team conducts research in areas integrating reinforcement learning and graph neural networks
The core area of use for the concepts we are developing is the energy system, in particular its design, control and automated energy trading.
We analyse trading in the cyclical market as a sequential decision-making process. An entity participating in a trade issues buy and sell orders. These orders should be determined on the basis of all information which shapes the current market situation and the situation of the trading entity. How to transform this information into orders in a manner that will bring the highest benefit to the entity? We design reinforcement learning algorithms that optimize strategies for determining orders based on market data.
The team develops neural networks that generate graphs. These solutions are oriented towards the automatic design of structures naturally represented by graphs, such as molecules or energy networks. Known methods for generating graphs are based on an assumed limitation on the size of the graph and are thus not scalable. How to generate arbitrarily large graphs that meet set functional requirements? We are developing methods that address this challenge.
While combining reinforcement learning and graph neural networks, we are working on reinforcement learning methods for multiple agents connected by graph-structured links. These methods are oriented towards optimising the control of devices in a system such as a power grid, such as distributed energy storage. We are looking for answers to the challenges of the ongoing energy transformation. More and more electrical energy is coming from weather-dependent sources, i.e. windmills and solar panels, which is desirable due to the promotion of renewable energy, but significantly impedes the controlling of the grid.