DARPA ULTRALOG Final Report - Industrial and Manufacturing ...
DARPA ULTRALOG Final Report - Industrial and Manufacturing ...
DARPA ULTRALOG Final Report - Industrial and Manufacturing ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Underst<strong>and</strong>ing Agent Societies Using Distributed Monitoring <strong>and</strong> Profiling<br />
† Wilbur Peng, † Vikram Manikonda, <strong>and</strong> ‡ Soundar Kumara<br />
† Intelligent Automation Incorporated<br />
7519 St<strong>and</strong>ish Place, Suite 200, Rockville, MD 20855<br />
{wpeng ,vikram}@i-a-i.com<br />
‡<br />
<strong>Industrial</strong> <strong>and</strong> <strong>Manufacturing</strong> Engineering<br />
310 Leonhard Building, The Pennsylvania State University, University Park, PA 16802<br />
{skumara}@psu.edu<br />
Abstract<br />
In this paper, we describe methodologies for<br />
underst<strong>and</strong>ing large-scale agent societies using the<br />
Castellan, a distributed profiling <strong>and</strong> logging system<br />
developed for Cougaar. Castellan enables the detailed<br />
efficient logging of blackboard plan activity. We describe<br />
the design, functionality, use <strong>and</strong> a number of<br />
applications of the Castellan tool, including a<br />
visualization <strong>and</strong> data mining tool based on a flexible<br />
algorithm for finding subgraph isomorphisms. By<br />
mapping “equivalent” meaningful graph nodes <strong>and</strong> edges<br />
to representative subgraph elements, the graph reduction<br />
approach reduces large plan graphs of hundreds of<br />
thous<strong>and</strong>s to millions of nodes to meaningful <strong>and</strong><br />
underst<strong>and</strong>able clusters <strong>and</strong> graph nodes. This algorithm<br />
is demonstrated through its application to event traces<br />
obtained from running Castellan within a military<br />
logistics planning society. In addition to providing data<br />
for static analysis after planning <strong>and</strong> execution, the<br />
Castellan approach is also useful for on-line analysis of<br />
active, running agent systems. We also describe a<br />
number of other potential applications of distributed<br />
monitoring for modeling, control, load balancing <strong>and</strong><br />
analysis.<br />
1. Introduction<br />
Distributed agent systems provide significant challenges<br />
for debugging, testing, profiling <strong>and</strong> tuning. Agent<br />
societies consist of distributed, state encapsulated entities<br />
that can run concurrently. Additionally, they have the<br />
additional constraint of state encapsulation, i.e. each<br />
agent does not have direct access to the state of other<br />
agents. Instead, they interact solely through message<br />
passing. Within an agent, different functionscan interact<br />
through sharing state.<br />
The Cougaar agent infrastructure supports an approach to<br />
distributed planning in which tasks are created <strong>and</strong><br />
exp<strong>and</strong>ed into subtasks by agents which can in turn be<br />
forwarded to other agents. The planning process creates a<br />
plan graph that spans multiple agents that can potentially<br />
be very large, growing to hundreds thous<strong>and</strong>s to millions<br />
of elements. Adding to the complexity of underst<strong>and</strong>ing<br />
system function, the plan graph generated by the agents<br />
can be dynamically modified during the planning <strong>and</strong><br />
execution phases of the society. As Cougaar agent<br />
societies increase in size <strong>and</strong> scope, underst<strong>and</strong>ing the<br />
distributed execution of the system becomes increasingly<br />
difficult. Being able to trace the time-evolving, eventdriven<br />
behavior across agents running societies becomes<br />
increasingly important.<br />
In this paper, we discuss methods for underst<strong>and</strong>ing,<br />
analyzing <strong>and</strong> controlling Cougaar agent societies<br />
through distributed profiling. Section 1.1 covers<br />
background concepts in distributed planning used by<br />
Cougaar. In Section 2, the Castellan profiling <strong>and</strong> logging<br />
system is introduced <strong>and</strong> its implementation <strong>and</strong> design<br />
described. Section 3 presents in detail an application of<br />
Castellan to data mining <strong>and</strong> visualization application<br />
using a plan graph reduction algorithm. <strong>Final</strong>ly, Section<br />
4 discusses potential applications of Castellan.<br />
1.1 Distributed plan graphs in Cougaar societies<br />
In this section, we review some basic concepts of<br />
planning in the Cougaar context. Additional details about<br />
plan representation can be found in [3].<br />
In Cougaar applications such as logistics planning <strong>and</strong><br />
execution, agents generate plans by decomposing tasks<br />
into subtasks, aggregating tasks, <strong>and</strong> forwarding tasks to<br />
other organization entities which are in turn are<br />
represented by other agents.