Elth Ogston

A method for decentralized clustering in large multi-agent systems

This paper examines a method of clustering within a fully decentralized multi-agent system. Our goal is to group agents with similar objectives or data, as is done in traditional clustering. However, we add the additional constraint that agents must remain in place on a network, instead of first being collected into a centralized database. To do this we connect agents in a random network and have them search in a peer-to-peer fashion for other similar agents. We thus aim to tackle the basic clustering problem on an Internet scale and create a method by which agents themselves can be grouped, forming coalitions. In order to investigate the feasibility of a decentralized approach, this paper presents a number of simulation experiments involving agents representing two-dimensional points. A comparison between our methods clustering ability and that of the k-means clustering algorithm is presented. Generated data sets containing 2,500 to 160,000 points (agents) grouped in 25 to 1,600 clusters are examined. Results show that our decentralized agent method produces a better clustering than the centralized k-means algorithm, quickly placing 95% to 99% of points correctly. The the time required to find a clustering depends on the quality of solution required; a fairly good solution is quickly converged on, and then slowly improved. Overall, our experiments indicate that the time to find a particular quality of solution increases less than linearly with the number of agents.

ARRG: real-world gossiping

Gossiping is an effective way of disseminating information in large dynamic systems. Until now, most gossiping algorithms have been designed and evaluated using simulations. However, these algorithms often cannot cope with several real-world problems that tend to be overlooked in simulations, such as node failures, message loss, non-atomicity ofinformation exchange, and firewalls.n We explore the problems in designing and applying gossiping algorithms in real systems. Next to identifying the most prominent real-world problems and their current solutions, we introduce Actualized Robust Random Gossiping (ARRG), an algorithm specifically designed to take all of these real-world problems into account simultaneously. To address network connectivity problems such as firewalls we introduce a novel technique, the Fallback Cache. This cache can be applied to existing gossiping algorithms to improve their resilience against connectivity problems.n We introduce a new metric, Perceived Network Size to measure a gossiping algorithms effectiveness. In contrast to existing metrics, our new metric does not require global knowledge. Evaluation of ARRG and the Fallback Cache in a number of realistic scenarios shows that the proposed techniques significantly improve the performance of gossiping algorithms in networks with limited connectivity. Even in pathological situations, with 50% message loss and with 80% of the nodes behind a NAT, ARRG continues to work well. Existing algorithms fail in these circumstances.

Group formation among peer-to-peer agents: learning group characteristics

This paper examines the decentralized formation of groups within a peer-to-peer multi-agent system. More specifically, it frames group formation as a clustering problem, and examines how to determine cluster characteristics such as area and density in the absence of information about the entire data set, such as the number of points, the number of clusters, or the maximum distance between points, that are available to centralized clustering algorithms. We develop a method in which agents individually search for other agents with similar characteristics in a peer-to-peer manner. These agents group into small centrally controlled clusters which learn cluster parameters by examining and improving their internal composition over time. We show through simulation that this method allows us to find clusters of a wide variety of sizes without adjusting agent parameters.

Collaborative filtering using random neighbours in peer-to-peer networks

Traditionally, collaborative filtering (CF) algorithms used for recommendation operate on complete knowledge. This makes these algorithms hard to employ in a decentralized context where not all users ratings can be available at all locations. In this paper we investigate how the well-known neighbourhood-based CF algorithm by Herlocker et al. operates on partial knowledge; that is, how many similar users does the algorithm actually need to produce good recommendations for a given user, and how similar must those users be. We show for the popular MovieLens 1,000,000 and Jester datasets that sufficiently good recommendations can be made based on the ratings of a neighbourhood consisting of a relatively small number of randomly selected users.

Group Formation Among Decentralized Autonomous Agents

This paper examines a method of clustering within a fully decentralized multi-agent system. Our goal is to group agents with similar objectives or data, as is done in traditional clustering. However, we add the additional constraint that agents must remain in place on a network, instead of first being collected into a centralized database. To do this, we connect agents in a random overlay network and have them search in a peer-to-peer fashion for other similar agents. We thus aim to tackle the basic clustering problem on an Internet scale, and create a method by which agents themselves can be grouped, forming coalitions. In order to investigate the feasibility of this decentralized approach, this paper presents simulation experiments that look into the quality of the clusters discovered. First, the clusters found by the agent method are compared to those created by k-means clustering for two-dimensional spatial data points. Results show that the decentralized agent method produces a better clustering than the centralized k-means algorithm, placing 95% to 99% of points correctly. A further experiment explores how agents can be used to cluster a straightforward text document set, demonstrating that agents can discover clusters and keywords that are reasonable estimates of those identified by the central word vector space approach.

The future of energy markets and the challenge of decentralized self-management

Complex, intelligent, distributed systems in dynamic environments, such as the power grid need to be designed to adapt autonomously. Self-management, in particular of large scale adaptive systems such as the power grid, is necessarily distributed. Agent and peer-to-peer based decentralized self-management can change the future of energy markets in which the power grid plays a core role. n nAssuming that both consumers and providers of energy are autonomous systems, represented by software agents or peers capable of self-management, virtual organizations of systems can emerge and adapt when necessary. Communication structures between systems, e.g., hierarchical or clustered organizations, can emerge, organizations between and within which systems can choose to cooperate and coordinate their actions, or compete. Overlay structures (as defined within p2p research) define such adaptive communication structures, multi-agent research provides interaction patterns. Global goals are achieved by local management on the basis of local goals and knowledge. The appropriate delegation of managerial responsibility determines the control structure. Aggregate information differs depending on the position of a system in an organization, the aggregation mechanisms and policies.

Peer-to-peer aggregation techniques dissected

Aggregation is the process of gathering and combining information from a number of sources. In peer-to-peer systems, aggregation is a basic component of a range of applications, including monitoring and complex-query resolution. Peer-to-peer aggregation services themselves are dependent on a number of other fundamental peer-to-peer services – directories, multicasting and system-size approximation. The overall performance characteristics of an aggregation service are affected by the chosen implementation method for these underlying services. To illustrate this relationship, aggregation techniques for internet-based peer-to-peer systems are surveyed and dissected into their component parts. We further consider the problem of running one-off aggregation queries in a peer-to-peer network. A new aggregation service, Bliksum, which uses a novel combination of underlying services, is introduced. Bliksum employs unstructured peer-to-peer techniques for node sampling, multicasting and system-size approximation, in combination with a method of building a temporary tree structure for aggregation itself. Unstructured peer-to-peer techniques have been shown to be highly resilient to node churn, avoiding the problem inherent in structured systems of maintaining the desired structure when the set of nodes changes rapidly. We present experiments showing that Bliksum retains these advantages while reducing communications cost and reducing information loss compared to pure gossip-based aggregation.

On the value of random opinions in decentralized recommendation

As the amount of information available to users continues to grow, filtering wanted items from unwanted ones becomes a dominant task. To this end, various collaborative-filtering techniques have been developed in which the ratings of items by other users form the basis for recommending items that could be of interest for a specific person. These techniques are based on the assumption that having ratings from similar users improves the quality of recommendation. For decentralized systems, such as peer-to-peer networks, it is generally impossible to get ratings from all users. For this reason, research has focused on finding the best set of peers for recommending items for a specific person. In this paper, we analyze to what extent the selection of such a set influences the quality of recommendation. Our findings are based on an extensive experimental evaluation of the MovieLens data set applied to recommending movies. We find that, in general, a random selection of peers gives surprisingly good recommendations in comparison to very similar peers that must be discovered using expensive search techniques. Our study suggests that simple decentralized recommendation techniques can do sufficiently well in comparison to these expensive solutions.

Peer sampling with improved accuracy

Node sampling services provide peers in a peer-to-peer system with a source of randomly chosen addresses of other nodes. Ideally, samples should be independent and uniform. The restrictions of a distributed environment, however, introduce various dependancies between samples. We review gossip-based sampling protocols proposed in previous work, and identify sources of inaccuracy. These include replicating the items from which samples are drawn, and imprecise management of the process of refreshing items. Based on this analysis, we propose a new protocol, Eddy, which aims to minimize temporal and spatial dependancies between samples. We demonstrate, through extensive simulation experiments, that these changes lead to an improved sampling service. Eddy maintains a balanced distribution of items representing active system nodes, even in the face of realistic levels of message loss and node churn. As a result, it behaves more like a centralized random number generator than previous protocols. We demonstrate this by showing that using Eddy improves the accuracy of a simple algorithm that uses random samples to estimate the size of a peer-to-peer network.

Improving the Accuracy of Peer-to-Peer Sampling Services

Node sampling services provide peers in a peer-to-peer system with a source of randomly chosen addresses of other nodes. Ideally, samples should be independent and uniform. The restrictions of a distributed environment, however, introduce various dependancies between samples. We review gossip-based sampling protocols proposed in previous work, and identify sources of inaccuracy. These include replicating the items from which samples are drawn, and imprecise management of the process of refreshing items. Based on this analysis, we propose a new protocol, Eddy, which seeks to minimize temporal and spatial dependancies between samples. We demonstrate that these changes lead to a better sampling service by showing, through simulations, that using Eddy improves the accuracy of a network-size estimation algorithm that uses the random samples from the protocol.