A simulator for a two layer MAS adaptation in P2P networks

Jordi Campos Miralles, Marc Esteva, Maite López-Sánchez, Javier MoralesWAT 2009

Abstract

This paper presents a specialized simulation environment for the Two-Level Assisted Multi-Agent System (2-LAMA) architecture. Designed to optimize Peer-to-Peer (P2P) networks, 2-LAMA introduces a distributed meta-level of assistant agents that dynamically adjust network norms and social structures. The simulator, built on the Repast Simphony framework, integrates high-level agent logic with a detailed packet-switching network model (netsim). This dual-level approach allows researchers to evaluate organizational adaptation strategies while accounting for physical network constraints such as bandwidth and latency.

Introduction

Multi-Agent Systems (MAS) rely on organizational structures—comprising rules, roles, and protocols—to coordinate agent behavior. However, static organizations often fail to maintain efficiency in the face of environmental volatility, such as fluctuating network congestion or shifting peer populations. Traditional solutions often involve increasing the internal complexity of individual agents to enable self-adaptation.

In open systems like P2P networks, where participants are heterogeneous and their internal logic cannot be globally controlled, this approach is often unfeasible. The authors propose the Two-Level Assisted MAS Architecture (2-LAMA), which offloads the burden of adaptation to a dedicated infrastructure layer. The primary challenge addressed in this work is the creation of a simulation environment capable of modeling both the high-level social logic of agents and the low-level physics of the underlying network.

The 2-LAMA Architecture

The 2-LAMA architecture bifurcates the system into two distinct layers:

  1. Domain Level (DL): This layer consists of the "peers" responsible for the primary task, such as data sharing. These agents can remain relatively simple, focusing on task execution rather than global optimization.
  2. Meta-Level (ML): This layer is composed of "assistant" agents. These assistants observe peer behavior, organize them into clusters, and share information with other assistants to optimize the system's global performance.
Figure 2: The 2-LAMA architecture showing the interaction between Meta-Level assistants (A1-A3) and Domain-Level peer clusters (P1-P12).

Unlike centralized "trackers," the meta-level is distributed. Assistants are ideally located at the ISP level, ensuring fault tolerance and scalability. They provide "Coordination Support" by offering advice and estimations to peers, helping them navigate the network more efficiently without mandating specific internal logic.

Simulation and Network Modeling

A core contribution of this research is the development of a dual-level simulator. While most MAS simulators treat message passing as abstract or instantaneous, the 2-LAMA simulator includes a custom netsim module. This module models a realistic packet-switching network where messages are decomposed into packets that traverse routers and links with finite bandwidths.

Figure 3: Layered view of the simulator, mapping the Meta-Level (ML) and Domain-Level (DL) logic to the physical Network topology.

This network-awareness is crucial because P2P performance is heavily influenced by physical constraints. The simulator allows researchers to visualize how agent-level decisions—such as neighbor selection—impact network-level metrics like link saturation and mean message hops.

Key Insights and Results

The 2-LAMA approach emphasizes the separation of concerns. By moving adaptation to the meta-level, the system can support "dumb" or heterogeneous agents while still achieving sophisticated global behavior. Furthermore, because the meta-level is distributed, the assistants must reach a consensus on system-wide "norms" (e.g., maximum allowed bandwidth or connection limits) using voting mechanisms.

Preliminary evaluations compared 2-LAMA against a standard BitTorrent-like protocol. The results indicated that 2-LAMA, through its dynamic social structure and norm adaptation, could identify more efficient configurations in specific data-sharing scenarios. The simulator's GUI provides real-time layouts of the P2P network, allowing for the immediate identification of bottlenecks and data distribution patterns.

Conclusion and Future Work

The 2-LAMA simulator provides a robust platform for testing organizational adaptation in P2P networks. By bridging the gap between social MAS logic and physical network simulation, it offers a more realistic assessment of coordination strategies.

Future research will focus on several key areas, including resilience to norm violators (freeloaders), modeling network churn as peers join and leave, and integrating machine learning so assistants can predict optimal norm adaptations based on historical data.

Summary updated

BibTeX
@inproceedings{miralles2009simulator,
  author = {Jordi Campos Miralles and Marc Esteva and Maite López-Sánchez and Javier Morales},
  title = {{A simulator for a two layer MAS adaptation in P2P networks}},
  booktitle = {WAT},
  year = {2009},
}
Figure