Haksun Li, Edmund Durfee, and Kang Shin
This paper studies the causes of over-utilization of the resource capacities of a group of cooperative resource-limited agents and what they can do to reduce their resource consumption. We view the problem of how an agent decides what tasks to execute and what to ignore as a type of multiagent planning problem. An agent in a multiagent setting has to prepare for all states it may reach as a result of its own actions, the environment transitions, as well as the actions potentially executed by other agents. Intuitively, the more information it knows about the plans of the other agents, the better it can allocate its resources for various tasks. Indeed, in our experiments over a particular sample space, on average, 50% of the actions are planned for the states it will never reach when an agent is completely ignorant about the plans of others. We have developed a protocol to allow agents to efficiently find out the relevant information about the plans of others. As the agent increases its level of awareness of the others’ plans, it can better identify the unreachable states to avoid spending resources on them.