Agent-based Distributed Resource Allocation in Continuous Dynamic Systems

Intelligent agents and multiagent systems reveal new strategies to design highly flexible automation systems. There are first promising industrial applications of multiagent systems for the control of manufacturing, logistics, traffic or multi-robot systems. One reason for the success of most of these applications is their nature as some form of a distributed resource allocation problem which can be addressed very well by multiagent systems. Resource allocation problems solved by agents can be further categorized into static or dynamic problems. In static problems, the allocations do not depend on time and many resource allocation problem of practical interest can be solved using these static considerations, even in discrete-event systems like manufacturing or logistic systems. However, problems especially in highly dynamic environments cannot be addressed by this pure static approach since the allocations, i.e. the decision variables, depend on time and previous states of the considered system. These problems are hardly considered in the relevant agent literature and if, most often only discrete-event systems are considered.

This work focuses on agent-based distributed dynamic resource allocation problems especially in continuous production systems or other continuous systems. Based on the current states of the distributed dynamic system, continuous-time allocation trajectories must be computed in real-time. Designing multiagent systems for distributed resource allocation mainly comprises the design of the local capabilities of the single agents and the interaction mechanisms that makes them find the best or at least a feasible allocation without any central control. In this work, the agents are designed as two-level entities: while the low-level functions are responsible for the real-time allocation of the resources in the form of closed-loop feedback control, the high-level functionalities realize the deliberative capabilities such as long-term planning and negotiation of the resource allocations. Herein, the resource allocation problem is considered as a distributed optimization problem under certain constraints. The agents play the role of local optimizers which then have to coordinate their local solutions to an overall consistent solution.

It is shown in this contribution that the described approach can be interpreted as a market-based allocation scheme based on balancing of supply and demand of the resources using a virtual price. However, the agents calculate and negotiate complete supply and demand trajectories using model-based predictions which also leads to the calculation of a price trajectory. This novel approach does not only consider the dynamic behaviour of the distributed system but also combines control tasks and resource allocation in a very consistent way. The approach is demonstrated using two practical applications: a heating system and an industrial sugar extraction process.