How can agents self-organize to share limited resources efficiently?
Self-organisation of common good usage and an application to Internet services
This paper explores how individuals choose between multiple shared resources ("common goods") when each individual's usage diminishes the resource's quality for everyone. It proposes a simple "Win-Stay, Lose-Shift" (WSLS) strategy: stick with a resource as long as it's satisfactory; switch randomly when it's not. This leads to surprisingly good distribution of usage. Introducing "selective" agents who adapt their tolerance for poor quality further improves the overall system, approaching optimal distribution. A key application is server selection in mobile networks, where users can choose between servers with different capacities and delays. Simulations validate the effectiveness of WSLS and adaptive tolerance in optimizing resource usage, even under fluctuating demand.
For LLM-based multi-agent systems, this work suggests: 1) Simple, localized strategies like WSLS can be effective for coordinating resource usage without complex communication or global knowledge. 2) Adaptability (like selective tolerance) further improves efficiency and robustness. 3) These concepts can be applied to managing computational resources in LLM-based multi-agent applications, potentially optimizing performance and cost by dynamically distributing workloads across available compute servers or LLMs with different capabilities.