Multi-agent Learning by Trial and Error for Resource Leveling during Multi-Project (Re)scheduling

Abstract

In a multi-project context within enterprise networks, reaching feasible solutions to the (re)scheduling problem represents a major challenge, mainly when scarce resources are shared among projects. The multi-project (re)scheduling must achieve the most efficient possible resource usage without increasing the prescribed project constraints, considering the Resource Leveling Problem (RLP), whose objective is to level the consumption of resources shared in order to minimize their idle times and to avoid overallocation conflicts. In this work, a multi-agent solution that allows solving the Resource Constrained Multi-project Scheduling Problem (RCMPSP) and the Resource Investment Problem is extended to incorporate indicators on agents’ payoff functions to address the Resource Leveling Problem in a decentralized and autonomous way, through decoupled rules based on Trial-and-Error approach. The proposed agent-based simulation model is tested through a set of project instances that vary in their structure, parameters, number of resources shared, etc. Results obtained are assessed through different scheduling goals, such as project total duration, project total cost and leveling resource usage. Our results are far better compared to the ones obtained with alternative approaches. This proposal shows that the interacting agents that implement decoupled learning rules find a solution which can be understood as a Nash equilibrium.