A multi-agent system for minimizing information indeterminacy within information fusion scenarios in peer-to-peer networks with limited resources

Highlights

• Heterogeneous peer-to-peer networks often depend on limited resources, which usually cuts down the performance of the network.

• We present a multi-agent information fusion system that relies on collaborating peers to significantly improve the quality of information in resource-limited settings.

• The underlying model is founded on querying the peers that have historically performed better for a given agent and information type.

• Our system has a broad spectrum of application domains, ranging from mobile recommendation systems to decision-making applications in critical environments.

Abstract

Information fusion (IF) has gained ground in recent years. It is increasingly used in applications involving networks of heterogeneous elements that communicate with each other peer-to-peer. This is thanks primarily to the advance of the Internet of Things (IoT) and the emergence of paradigms like holonic information fusion. However, heterogeneous peer-to-peer networks often depend on limited resources (energy, communication capacity, time, etc.). On these grounds, network components necessarily have to behave intelligently. They also have to be autonomous and be able to coordinate their actions in order to obtain results in the presence of vague or uncertain data. In this paper, we present a multi-agent information fusion system model that relies on collaborative peers to improve the quality of the information handled by the agents. The idea behind the model is to query the peers that have historically performed better for a given agent and information type (such as a certain data field). We report the results of the experiments we conducted on a proof-of-concept implementation of the proposed system model, consisting of a statistically significant number of simulation runs on two case studies with different numbers of agents and messages. The results show that the performance of an open peer-to-peer network of agents with no predefined structure, measured as mean traffic per agent (i.e. the use of resources) for replies of different quality and as the strict success rate, improves significantly when the members of the network adopt the intelligent mechanisms proposed in this article. This system model has a broad spectrum of application domains, ranging from mobile recommendation systems to decision-making applications in critical environments.

Introduction

The sustained increase in systems connectivity, the advent of the Internet of Things (IoT) and the emergence of the holonic information fusion (IF) paradigm [42], [34] have led to the fusion of more and more information coming from heterogeneous data sources. These data sources communicate non-hierarchically through a communications network and their resources (energy, quantity of transferable data, etc.) and information quality (which is vague, uncertain, incomplete, etc.) are limited.

In this regard, this paper focuses on describing a system model whose components fuse information. The aim of IF is to improve the quality of the information that an agent querying the network receives from its peers. All (peer-to-peer) interactions are possible, although they are gradually restricted over time for the purpose of selecting the most reliable peers with a view to saving resources in limited resources scenarios.

IF is carried out in order to reduce the negative impact of information imperfections [19], [42] (such as the vagueness and uncertainty of the received messages) on decision-making. Impact reduction means that the receiver should make the best decisions in view of the semantic content assigned to the received message. Indeterminacy is a key feature of information. Uncertainty and vagueness are two basic aspects of indeterminacy. Uncertainty is primarily associated with a data error or imprecision, and vagueness is typical of natural language (in the phrase “lengthy paper”, for example, how many pages means “lengthy”?). On the other hand, information quality usually denotes a multidimensional characteristic of information. Information quality is composed of many facets (also referred to as characteristics or dimensions), which are still open to debate. Of these facets, we are concerned here with vagueness and uncertainty: how accurate or reliable a data item is and how fuzzy this data item may be (see Section 2.1). In this regard, we compare two quality measures by synthesizing their dimensions (vagueness, uncertainty, and others related to reliability, credibility, and efficiency) in order to form a value that is called quality (see Section 2.2).

The main contribution of this paper is the proposal of an original multi-agent system model based on a heterogeneous peer-to-peer network with limited resources (like time, energy, number of messages, processing capacity, bandwidth, etc.), whose human or other agents fuse information to reduce the chances of poor quality information. This model's key innovative feature is that it is capable of switching from an unrestricted to an intelligent operating mode when agents detect a situation (event) where certain conditions apply, for instance, a number of unanswered messages within any preset time limit or a specified limit on the number of available messages per agent. In this manner, it can optimize the use of resources to achieve best system performance. It is this feature that distinguishes our investigation from the conventional research focus on maximizing the quality of the transmitted and received information regardless of limited resources of individual system agents. In intelligent mode, each system component employs decentralized information fusion, using local data either competitively or cooperatively to increase the quality of the information contained in the indeterminacy messages that it exchanges with and receives from the queried agent, focusing on vagueness and uncertainty, in order to make the best decision. The model is based on an information quality metric. Our proposed metric takes into account information uncertainty and vagueness typical of human communication. Being scalar, this metric supports decision-making and automation.

Our model employs event-triggered transmission schemes that have previously been proved to be useful for saving resources in networked systems management [20], [21], [51]. While the reported systems are based on resource allocation and nodes communication assignment strategies, our system uses an information fusion strategy based on querying the peers that have performed better for a given agent and information type (such as a certain data field) in the past.

Additionally, we provide empirical evidence on the use of JDL Model Level 4 information quality for information fusion process management [27], [45]. In this respect, the proposed model adaptively selects the data to be acquired by picking the better quality data from the responding (available) sources, whereas the queried sources change over time. This is a contribution to what Bossé et al. [42] described as the scant research on this topic.

The effectiveness of the proposed system was validated by means of an intensive simulation study, which demonstrated a statistically significant improvement in system performance and success rate compared with the same unintelligent system. System performance was measured as mean traffic per agent for different response qualities, and system success as agents behaving in accordance with the proposed model.

The implementation of the proposed model is potentially applicable in real networked systems. It has a broad spectrum of application domains, ranging from mobile recommendation tools to decision-making applications in critical environments and can be used in different fields including the IoT, human organizations, etc.

This article is structured as follows. Section 2 reports the research related to IF and peer-to-peer (P2P) networks. Section 3 formally states the problem. Section 4 describes the problem solution as a generic system architecture. Section 5 then presents and analyses the results of the experiments conducted on a simulation of a particular implementation and two case studies. Section 6 describes the model applicability. Finally, Section 7 outlines the conclusions and future research.