D'Agents and CoABS
DARPA's CoABS ("Control of Agent-Based Systems") project is a major sponsor of the D'Agents project at Dartmouth College.
Mobile agents, which are programs that can migrate under their own control from host to host in a network, have the potential to radically change how distributed information-gathering applications are constructed. By moving to an appropriate network location, a mobile agent can avoid excessive use of low bandwidth, high latency or unreliable network links. If (1) the status of the links changes, (2) new information sources become available, or (3) existing information sources go offline (forcing the agent to find and use different sources), the agent can reposition itself in response to the new conditions. Mobile agents are particularly attractive in mobile computing, since the user often has a low-powered device and often is disconnected from the network. A mobile agent can migrate off of the low-powered device so that it can run on a higher-powered machine and continue its task even if the network link goes down.
Although mobile-agents have not yet realized their full potential, significant strides have been made in recent years on performance, security and fault tolerance, and mobile agents are now an effective choice for many information-retrieval applications. One remaining issue, however, is resource allocation. Resource allocation is a critical component in any mobile-agent application, particularly complex information-retrieval applications. All resources must be effectively utilized, load must be balanced evenly, and agents must avoid chaotic or overly consumptive behavior that can lead to ineffective applications and systems. Resource allocation is particularly difficult in large-scale systems that span multiple administrative domains and include legacy applications as subsystems.
Human social and economic systems have long encountered similar problems involving resource allocation and productive cooperation among autonomous agents. Decentralized market-based systems evolved to allow the overall economic system to operate efficiently, suppressing chaotic and resource-wasting behaviors. Our project will design, study and implement similar market-based mechanisms for distributed computer systems built around mobile agents. The main scientific challenges in this work involve the combination of relatively well-understood economic concepts with specific computer-science issues involving agent design, distributed directories, on-line routing, secure communication, and fault tolerance.
We anticipate a future in which mobile agents are widely used in military, commercial, and educational distributed systems, made up of computing devices from palm-sized organizers to scalable clusters of workstations to the Internet itself, connected by a variety of networks from wireless modems to Fast LANs to wide-area internets. Information resources will be constantly added, updated, and moved. New applications for these information resources, applications never anticipated by the resource creators, can be prototyped and installed easily through mobile-agent technology. Our method of global resource control will ensure that all participants have fair and effective access to the resources, and will prevent bottlenecks, system failures, resource-consumptive attacks, and chaotic behavior.
In particular, we will use a market-based model of global resource control, in which virtual currency is used to control the global resource usage of a mobile agent, and in which information providers/servers control demand by adjusting by adjusting the price of their service. Virtual currency provides a common language for agents to evaluate the cost of various alternatives, such as the load on multiple servers or the latency of various network links. In the proposal, we show how a market-based model can be a unified approach to many critical problems in resource control, including load balancing, agent planning, agent navigation, and the avoidance of consumptive behavior. Furthermore, market-based control is extremely robust due to its decentralized control, and provides a clean solution to security problems like denial-of-service attacks.
We will demonstrate an agent infrastructure that can support a wide range of computing devices and networks, and that can scale efficiently to thousands of agents running on workstation clusters that are tightly coupled in a LAN or on clusters that are scattered around the globe. We will also demonstrate that market-based control can be a unifying approach to the effective solution of many of the associated resource-control problems.