Distributed Multiagent Resource Allocation with Adaptive Preemption

We propose and validate a novel approach to multiagent resource allocation, with special attention to supporting resource preemption in a fully distributed, cooperative setting. Our solution utilizes planning and learning techniques to obtain reasoned local estimates of the impact of preemptions on the system as a whole. Agents are thus able to effectively coordinate the allocation of resources on large scales and produce efficient, high throughput solutions to resource allocation problems, allowing important preemptions for improved utility. We perform simulations of multiagent system environments with uniform agent type distributions and limited numbers of resources to ensure that preemption requests will occur. We then demonstrate that our system produces effective global utility, measured in terms of the average costs and wait-times incurred by agents in need of resources, as well as improvements to a standard baseline.