Distributed agents for cost-effective monitoring of critical success factors

Abstract

Business managers should promptly respond to important events (e.g. exceptions) that happen on a set of critical success factors (CSF). A CSF monitoring system is thus essential in capturing the events for the managers. It monitors the information items concerning the CSF. Once an update is detected, critical events may be validated, logged, and signaled for the manager. Since CSF monitoring is often time-critical and mission-critical, a CSF monitoring system should be cost-effective: It should detect updates in a timely, complete, and robust manner without incurring heavy loading to related information servers (e.g. query overheads) and the Intranet (e.g. communication overheads). To achieve that, the monitoring tasks should be properly distributed and coordinated on the Intranet. We propose a multiagent CSF monitoring paradigm, CSFMonitor, in which distributed agents share a collective goal of cost-effective CSF monitoring. An experiment on monitoring real-world financial CSF is conducted. The delivery of CSFMonitor to businesses may robustly provide more complete and timelier information without causing serious problems to the original information processing in businesses.

Introduction

Management by exceptions (MBE) has been widely adopted in business administration. It suggests managers to promptly respond to the exceptions happening on a predefined set of critical success factors (CSF), without being involved with the tedious activities of monitoring and validation [3], [16]. CSF monitoring is thus essential to the realization of MBE. Conceptually, a CSF consists of a critical information item and a validation procedure. The critical information item may be maintained by internal or external entities (e.g. database systems, web servers, and file servers) [14] and updated at any time [32]. Once an update on the information item is detected, the validation procedure is triggered to determine whether an exception occurs, and if so, corresponding managers are notified. By promptly responding to the exceptions happening on the CSF, the managers may maximize the wealth of their businesses.

In this paper, we explore how multiagent systems may support CSF monitoring so that managers may promptly respond to a higher percentage of exceptions. CSF monitoring is actually a kind of information monitoring, since exceptions may be detected only by monitoring the updates of their corresponding information items. However, since CSF monitoring is often more time-critical and mission-critical, it is associated with particular requirements on timely, complete, and robust (fault tolerant) update detection. That is, a CSF monitoring system should detect a higher percentage of updates in a timely manner without being crashed by any single fault or failure.

The requirements significantly increase the cost of CSF monitoring. We thus focus on the design of multiagent systems for cost-effective CSF monitoring, which aims to maximize the quality but minimize the cost of CSF monitoring. On the quality part, we are concerned with the timeliness, completeness, and robustness of update detection. On the cost part, we are concerned with the loading incurred by the monitor to the related information servers, the Intranet, and the Internet. The amount of the cost should be properly controlled; otherwise, the delivery of the system will make all the servers and the Intranet exhausted, which, in turn, deteriorates the performance of original information processing in businesses [23]. Similar situations may be found in various domains such as information inquiry through the Internet [21]. A cost-effective CSF monitor should maximize timeliness, completeness, and robustness of update detection, while simultaneously minimize the cost it incurs.

Major challenges of cost-effective CSF monitoring lie on the trade-off between the quality concerns and the cost concern. For example, timely and complete monitoring often calls for frequent inquiry to information servers, which will incur heavy loading (cost) to the servers and the Intranet. As another example, consider robust monitoring, which may be achieved by distributing the monitoring tasks (i.e. the tasks of monitoring individual CSF) to multiple machines. In that case, a fault on a machine will not cause serious problems to CSF monitoring. However, intensive communications among the distributed tasks should be avoided for not incurring a heavy burden to the Intranet.

Previous information monitoring techniques did not tackle the challenges, making them unsuitable for CSF monitoring. They often relied on a centralized site to monitor all information items [9], [24], [26], [27], [30]. This method cannot be robust, since a fault on the centralized site could stop all monitoring tasks, and thus cause a great loss to the managers. Moreover, previous techniques often predefined a frequency to periodically check for information updates. This method could not fulfill the requirements of CSF monitoring either, since critical information items may be updated at any time. They may even be updated frequently at a particular time (e.g. the daytime of some particular days in a month) but infrequently at another time. Since periodical monitoring does not estimate the timing of information updates, it may waste a great amount of efforts without finding information updates. On the other hand, the timing of information updates cannot be hypothesized by traditional forecasting techniques (e.g. moving average) either, since the monitoring system cannot know the happening time of the updates. (Note: Only the servers that maintain the information items may know the happening time. Unfortunately, in most cases, they cannot provide such information to the monitoring system.) No reliable history data may be sampled in order to forecast the happening time of the next update.

We tackle the challenges by delegating a monitoring task to an agent, which may be distributed to various sites on the Intranet. The distributed setting contributes three main benefits to CSF monitoring: load balancing, bandwidth saving, and robust monitoring. This is because the agents may be distributed according to the loading (for load balancing) and the location (for bandwidth saving) of each site. A fault on a site will not stop all the monitoring tasks (i.e. robust monitoring). Therefore, under such a distributed setting, we explore how individual agents may learn to monitor their CSF and collaborate with each other so that a limited amount of resources (i.e. query transactions and Intranet bandwidths) may be directed to those agents that are more likely to detect information updates. The amount of the communications required for the collaboration should be controlled as well, since intensive communications among the distributed agents may cause serious problems to both the efficiency of the collaboration and the loading of the Intranet.

We first survey the environments and requirements of distributed cost-effective CSF monitoring for decision support (ref. Section 2). Based on the requirements, a multiagent paradigm CSFMonitor is developed (ref. Section 3). Each agent learns to properly work with each other to achieve cost-effective CSF monitoring. Cost-effectiveness of CSFMonitor is investigated in an experiment that simulates real-world environments of financial CSF (ref. Section 4). In the experiment, CSFMonitor significantly outperforms state-of-the-art monitoring techniques. The framework is then comprehensively evaluated from the viewpoints of its related work and future work 5 Discussion, 6 Future work. We finally conclude that effective decision making may be supported by the delivery of CSFMonitor, which aims to robustly and promptly report more events happening on CSF without causing problems to original information processing in businesses.