Dynamic Plan Evaluation for Military Logistics

M. Afzal Upal and F. Fung (USA)


Automated military logistics planning, decision theory


The agent interactions are governed by the roles that the interacting agents are playing at the time. For instance, a superior agent may pass requirements (tasks and Oplan policies) down to its subordinates and the subordinate agents may pass the status back up to their superior. A customer agent, on the other hand, may request a service from its supplier (such as transport this for me, supply me with this) and a supplier may inform its customers about the status of their requests. The defense personnel engaged in logistics planning need tools that allow them to evaluate a plan and efficiently update this information as the situation evolves. This paper reports on the work that IET performed as part of DARPA's Ultra*Log project to design a probabilistic reasoning algorithm for dynamic plan evaluation that conforms to the stringent performance requirements of Ultra*Log's multiagent planning architecture called Cougaar. Cougaar users and agents specify their requirements in terms of tasks. Similar to classical AI planning tasks, a Cougaar task consists of a verb such as move and an object on which the verb operates such as Truck5931. However, a Cougaar task may also contain the preferences of the specifying agent about resolution of the task such as "I prefer those resolutions of this task that minimize execution time". Cougaar preferences may be specified as a value function over the Cougaar defined world aspects which include cost, duration, start time, and end time of a task. A Cougaar preference function is defined as a weighted sum of the aspect scoring functions.

Important Links:

Go Back