NewVectors Projects
Agents Overcoming Resource-Independent Scaling Threats (AORIST)
NewVectors’ Agents Overcoming Resource-Independent Scaling Threats (AORIST) enables the designers and users of agent-based resource management systems to detect and manage dynamic pathologies that emerge through the feedback between the changes agents make in the environment and decisions that agents base on the state of that environment.
In multi-agent systems, fundamental global system properties emerge as the aggregate effect of local decision strategies. Both global and local properties can be either exogenous (imposed by the programmer, system operator, nature or an adversary) or endogenous (generated by the agents). These distinctions imply four classes of information. Preliminary experiments with a simple model of resource allocation show that appropriate mathematical methods applied to these four categories of information can predict, detect, and control global system properties. Inter-agent endogenous information is generated by the agents themselves, so it reflects the decisions of individual agents even when those decisions are not individually accessible.
In turn, this information guides subsequent decisions. The resulting feedback can generate more information than the agents themselves can access. Thus inter-agent (global) endogenous information has great potential for system-level prediction and control, particularly when related to inter-agent information using techniques inspired by statistical physics.
This emergent global system behavior may be pathological, generating complex dynamics and instability that must be understood and controlled. NewVectors’ methods can characterize global behavior in ways that are vital to system performance, but not intuitively obvious. For example, agent deployment depends on understanding the relation between global dynamics and scale issues such as total agent population. Classical analysis of scaling and system performance focuses on the threat of a combinatorial explosion in the demands on computing and communication resources. The situation is more complex. We have identified a scaling threat that is independent of the demands agents place on system resources (thus, a "Resource-Independent Scaling Threat" or RIST). This threat compromises system performance when the population is either too large or too small compared with the sophistication of the agents.
If this dynamic is not taken into account, an agent prototype that performs acceptably on a test problem may under-perform when it is fielded on a larger or smaller scale. AORIST will provide the tools to characterize, predict, and control RIST and similar global dynamic behaviors of agent-based systems.
NewVectors’ AORIST is a project under the Agent Networking for Task Scheduling (ANTS) program. Two co-PIs reflect its interdisciplinary character, bridging agent technology and statistical physics. The PI for agents is Dr. H. Van Dyke Parunak, NewVectors' Chief Scientist. The PI for statistical physics is Professor Robert Savit, a member of the faculty in the University of Michigan's Physics Department, and founding director of the University's Center for the Study of Complex Systems.