Improving Multi-agent Based Scheduling by Neurodynamic Programming

Scheduling problems, e.g., a job-shop scheduling, are classical NP-hard problems. In the paper a two-level adaptation method is proposed to solve the scheduling problem in a dynamically changing and uncertain environment. It is applied to the heterarchical multi-agent architecture developed by Valckenaers et al. Their work is improved by applying machine learning techniques, such as: neurodynamic programming (reinforcement learning + neural networks) and simulated annealing. The paper focuses on manufacturing control, however, a lot of these ideas can be applied to other kinds of decision-making, as well.