Széchenyi Plusz RRF

In traffic applications, computational intelligence is gradually emerging as an indispensable component, playing a crucial role in enhancing system performance and decision making processes. Nowadays, the question is no longer whether tasks can be performed, but rather the rapid and energy-efficient production of agents, as well as their reusability. In traffic network management, static objects such as intersections can be treated as agents, which, by monitoring their local environment, make event-driven interventions at traffic signals. Since many traffic networks contain intersections with identical architectures, the homogeneous architecture of individual agents is a well utilizable property, as they can be represented by a common neural network, given that their observation and intervention spaces are identical. Although previous research has already shown that deep learning models trained on a network of a given size are capable of providing better solutions compared to other adaptive methods, even after altering the number of agents, the issue of which network is most suitable for performing the long and resource-intensive task of training remains unexplored. The aim of the research presented in the paper is to demonstrate that, during training on larger networks, agents learn about their environment in a more diverse manner, allowing them to better adapt to changes in the number of decision-makers.