Reinforcement learning for solving shortest-path and dynamic scheduling problems
Stefán Péter, László Monostori, Ferenc Erdélyi · SZTAKI Publication Repository (Hungarian Academy of Sciences) · 2001
Reinforcement learning (RL) has come to the focus of agent research as a potential tool for creating adaptive agents’ “learning module”. Its application is proved to be useful in cases where agents are performing high number of iterations with their dynamically changing environment. In the paper RL-based implementation of the Internet protocol (IP) packet routing problem is considered. A network of adaptive and cooperative agents are to transport data following the shortest route in time from a source computer to a destination, even if the transfer capacities of connections change in time. In contrast to the famously used static routing algorithms, the proposed method is capable to cope with dynamic conditions as well. Other possible applications of RL are also treated in the paper, including The Travelling Salesman’s problem, which implies solving dynamic scheduling tasks as well.