Visualizing multi-agent collaboration for classification of information



Agent-based simulation is a popular technology for studying social and information systems. Information visualization in such simulations is potentially useful for communicating real time information but involves several levels of challenges. It remains unclear what patterns of agent activities are useful for visualization. To motivate and investigate potentially useful patterns for visualization, we identified four factors: 1) Agent Involvement; 2) Dominant Player; 3) Learning and Adaptation; and 4) Influence of Task and Content Stream. To demonstrate the usefulness of the factors, we operationalized them to permit analysis and visualization in the context of a problem involving coordination among agents conducting document classification.


A multi-agent framework is useful for studying complex social and information systems. By definition, an agent is a computer program capable of autonomous action to meet its designed objectives in certain environment (Jennings & Wooldridge, 1998). In multi-agent systems, agents are treated as distributed peers that have scattered intelligence and can collaborate with each other to do certain tasks. Research on information retrieval has relied upon multi-agent technologies for better understanding of collective retrieval operations in distributed environments. This also responds to the increasing computational demands for retrieval and offers a great potential of scalability.

However, a common challenge in the use of agent technology involves the monitoring of agent activities. User interfaces should be designed to convey information about real-time dynamics in the systems. Information visualization, which attempts to facilitate information communication through visual representations (Hearst, 1999), is potentially useful in the agent-based research. Jennings and Wooldridge (1998) points out that visualization is particularly important for the development and use of agent technology because it is critical, though difficult, to know what is happening in asynchronous, concurrent systems. A few agent software frameworks have specialized features that support information visualization. However, it remains unclear what type of information is commonly useful for visualization and applicable across systems.

Proposed Pattern Factors for Agent Visualization

The objective of this paper is to propose factors that can be used to characterize important patterns of agent dynamics and to apply these factors to the analysis and visualization of existing agent systems. We introduce the following factors based on a review of common characteristics of agent communities for information retrieval.

1. Agent Involvement

In an agent society, agents have limited knowledge/intelligence individually and have to work with each other in order to accomplish given tasks. A common optimization objective is to involve as few agents as possible. The Agent Involvement factor denotes number of agents (or, computational resources) required to “solve” a particular task (e.g., to find the best class for an information item or a document).

2. Dominant Player

Agents may play different roles in their communities, especially in heterogeneous environments. Even they are designed to be homogeneous, their importance and the ways they connect and work with others can change over time. The Dominant Player factor helps us to identify agents that make major contributions to the local/global society, either because of critical resources they have or because of important gates/connections they maintain. We can detect which agents are most helpful in a community as they relate to a task and a particular set of information items.

3. Learning and Adapation

Learning and Adaptation is an important characteristic of all agent societies. Collective intelligence emerges from (even primitive) agents that collaborate with each other. Over time changes in collaboration patterns take place. For example, dominant players may emerge and and involvement in collaboration may decrease when agents have learned to collaborate effectively.

4. Influence of Task and Content Stream

Agent systems are designed to be adaptive and sensitive to changes. Changes of information items and/or tasks assigned to the agents will have influence on how they work with each other. To study the impact, one can systematically change how each information item is submitted to the agent community and determine how collaboration patterns change.

MACCI: A Collaborative Agent Community

We have developed a multi-agent simulation framework called MACCI, or Multi-Agent Collaboration for Classification of Information. The agents have limited/distributed knowledge for document classification1 and collaborate with each other to overcome the knowledge distribution. Each agent was equipped with a certain learning algorithm for predicting potential collaborators, or helping agents. A user interface with functionality to control and monitor the agent community through visualization has also been implemented, as shown in Figure 1 (a).

Figure 1.

MACCI: A Collaborative Agent Community

Figure 1 (b) illustrates how the multi-agent system works. The document distributor loads the document collection and releases one document to a randomly selected classifier agent each time. The chosen agent takes the incoming document and tries to classify it. If it fails to classify the document locally, it asks another agent for help by sending the document to the remote agent. If the remote agent succeeds in classification, it sends the result back. Otherwise, it asks another agent for help if only the range of collaboration is smaller than a pre-defined maximum number, and so forth. Only after the classification result is back will the document distributor release another document.

Visualization Results and Discussions

Research on MACCI has produced competitive retrieval results in the distributed manner (Fu, Ke, & Mostafa, 2005, Ke, Mostafa, & Fu, 2007). Based on previous experimental log files, we examined the agent collaboration activities and applied the proposed factors to produce the visualizations in Figure 2 (a) (b). Below, we discuss the visualizations.

Figure 2.

Agent Visualizations

Agent Involvement Over Time

Figure 2 (a) shows the evolution of Agent Involvement, or number of agents involved in each task, over time. Since the documents were streamed at the agent society one after another, we used sequential document ID to form the time axis (X). Each vertical bar represents a task (document classification in this case) and the height corresponds to the number of agents involved for the task. Black bars (or darker color on black/white print) denote successful2 collaborations; yellow (or gray on black/white print) means failure. The white curve going through the middle of the bars shows smoothed average Agent Involvement over time. The gradually decreasing trend of the curve indicates that less agents were involved in each task as they learned to collaborate more effectively. A closer look at the map also shows larger blocks of black bars near the end of the time axis, indicating more successful collaborations through learning. Dominant Players

Figure 2 (b) characterizes how the agents worked with each other and the dominant players in the community. On the map, a node denotes an agent while an arc (directed line) indicates the collaboration or communication path (line arrow, the direction of asking help). Node size depicts the frequency of engagement per agent. Node border color denotes the ratio of successful collaborations–the darker the color, the more successfully the agent had helped others. Line width (thickness) denotes the frequency of collaborations/communication between the two agents connected3; line color (darkness) indicates the ratio of successful collaborations–the darker, the higher percentage of successful collaborations.

By looking at node size and border color, we can easily identify several dominant players on the map, e.g., and Agent14Agent22. It is also obvious that these dominant players had intensive collaborations between each other. Interestingly, the visualization shows asymmetric connections among the agents–most of the connections are unidirectional. For example, Agent1 was very helpful to Agent12 as suggested by the thick and dark directed arc from Agent12 to Agent1. But Agent12 was NOT helpful to Agent1 as there is no directed arc back.

Summary and Future Work

In this paper, we introduced four pattern factors for the analysis and visualization of agent systems, namely, 1) Agent Involvement; 2) Dominant Player; 3) Learning and Adaptation; and 4) Influence of Task and Content Stream Although the visualization results were at a preliminary stage and they need to be combined with other types of analysis, some useful patterns were observed based on the four proposed factors. We are sufficiently encouraged to pursue the visualization approach for further investigation. Future research will involve examination of the factors in larger document collections and different tasks.


Authors acknowledge the NSF grant ENABLE #0333623.

  1. 1

    1 We define a classification task to be, given an information item (e.g., in this case a textual document), finding the best topical label or class.

  2. 2

    2 By successful, we mean one agent involved in collaboration could and did offer help.

  3. 3

    3 A threshold has been applied to the line width in order to show critical connections.