A Project of Dr. George Markowsky’s Homeland Security Research Lab

Under the direction of Dr. Thomas Wheeler

Background Information and the Problem

Karpagavalli Vel of Dr. Markowsky’s research lab is working on a system to identify possible terrorist threats, whether they be using aircraft as weapons, chemical or biological attacks, etc. She is currently building a knowledge base of facts gathered from newspapers and the Internet, which will be used with a system implemented in Prolog running on a single machine. This is a first prototype of a system to build up a starting knowledge base and to actually start making meaningful inferences about the knowledge. If such a system were to actually be deployed this system will have to deal with a vast amount of information in a reasonable amount of time.

Our solution to the performance issue introduced by the huge amount of data the system would have to deal with is to create a distributed system that would parallelize the inferences by having several Prolog programs running on different machines making inferences on a different set of data with different sets of rules. Each node will pass along any inferences to other nodes that may find the information useful. For example if one node has information that may be useful to a node making inferences about airplanes as weapons and another node infers that someone is a potential plane hijacker then that information should be passed along. New inferences received from other nodes will then be added as rules. New inferences will move their way up the tree where Prolog programs will be making inferences based on facts that were obtained as a result of inferences at a lower level.

Our initial plan is to use Python to handle communications between nodes. The choice to use Python was made because Dr. Wheeler has previous experience using a Python-Prolog combination. It is his experience that Python’s ease of network programming and Prolog’s ease of inferences makes the combination a good candidate for the implementation of our system. The Python programs would be responsible for communication between nodes and would run Prolog programs at each node. This seems like a good solution as Prolog is great for the inference engine, but is not suited well to the other aspects of the system.

Communication between nodes would be achieved through message passing, which is the natural interaction mechanism of Beowulf. Communication should be kept to a ‘need to know’ basis, as only useful inferences should be sent to other nodes. In a large distributed system that is processing a vast amount of information it is possible to overload the communications system so this has to be given consideration.

One interesting aspect to the project is that each node should be able to tell if any inference will be useful, and to what other node it would be useful to, rather than just sending any new inference to every other node.

My initial goal is to use the UMaine Computer Science Beowulf cluster to run the system. After each node can run small Prolog programs and can send inferences between each other we will begin moving the existing single-machine system to our parallel system, and make modifications as necessary.