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.
Manuscript for IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS 3<br />
generating global solution (i.e., completion time). For a given topology, the network can control<br />
its behavior by utilizing two different kinds of control actions: algorithm selection <strong>and</strong> resource<br />
allocation. While resource allocation tries to efficiently utilize limited resources, algorithm<br />
selection can change the amount of required resources. The resource allocation we are addressing<br />
here, is allocating resources of each machine to the residing components for a given topology. As<br />
problems are decomposed in various ways depending on their nature <strong>and</strong> size, <strong>and</strong> their QoS<br />
functions are context-dependent, the network needs to provide adaptive solutions to given<br />
problems by utilizing such control actions.<br />
One can imagine wide range of scientific <strong>and</strong> engineering problems that can be solved by<br />
such a network. UltraLog (http://www.ultralog.net) networks, implemented in Cougaar<br />
(Cognitive Agent Architecture: http://www.cougaar.org) developed by <strong>DARPA</strong> (Defense<br />
Advanced Research Project Agency), are the instances [5]-[9]. Each agent in these networks<br />
represents an organization of military supply chain <strong>and</strong> has a set of components specialized for<br />
each functionality (allocation, expansion, inventory management, etc) <strong>and</strong> class (ammunition,<br />
water, fuel, etc). The objective of an UltraLog network is to provide an appropriate logistics plan<br />
for a given military operational plan. A logistics plan is a global solution which is an aggregate<br />
of individual schedules built by components. An operational plan is decomposed into logistics<br />
requirements of each thread for each agent, <strong>and</strong> a requirement is further decomposed into root<br />
tasks (one task per day) for a designated component. As a result, a component can have hundreds<br />
of root tasks depending on the horizon of an operation <strong>and</strong> thous<strong>and</strong>s of tasks to process as the<br />
root tasks are propagated. As the scale of operation increases there can be thous<strong>and</strong>s of agents<br />
(tens of thous<strong>and</strong>s of components) in hundreds of machines working together to generate a<br />
logistics plan. QoS of these networks is determined by the quality of logistics plan (value of