Granular value-function approximation for road network traffic control

The research discussed in this paper aims at developing fast stable learning agents for large-scale complex systems including network traffic signal control systems. The control system is based on reinforcement learning (RL), an important research area in distributed AI with a wide area of applications including real-time control. RL-based control may also be suitable for distributed domains that are subject to time and environmental contingencies. Based on this assumption, the goal in this paper is to investigate ways to make RL excel at on-line, continuous state and action space tasks by incorporating the concept of fuzzy granulation as (powerful) function approximation tool: we argue why this may strongly improve the learning speed of the algorithm. The potential implications of this research are better running times, allowing us to consider much larger problem sizes.