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.
• The set of expansion edges X is subdivided into<br />
a set of disjoint subsets which have the same<br />
source node.<br />
• A task t is connected to exactly one task by an<br />
edge e unless e is a member of the set of<br />
expansion edges X.<br />
The basic principle behind the graph reduction approach<br />
is as follows. For each type of reduced graph mapping R,<br />
we define equivalence criteria between nodes in the graph<br />
to find abstract nodes. An example of criteria C 1 would<br />
be “All tasks at the same agent with parent tasks<br />
originating from another agent.” In applying the graph<br />
reduction algorithm, all nodes which satisfy this criteria<br />
are aggregated into a single node.<br />
Also, for every graph reduction mapping R, we define a<br />
equivalence criteria for abstract edges. Abstract edges<br />
are aggregates of equivalent subgraphs into a single<br />
representative edge.<br />
Continuing the example, consider a criteria C 2 that states<br />
“Subgraphs that connect all aggregate nodes satisfying C 1<br />
<strong>and</strong> connect to organizational assets associated with the<br />
same agent as C 1 .” Also, we define a second node<br />
equivalence criterion C 3 as “All tasks at the same agent<br />
which are allocated to assets representing external<br />
organizations.”<br />
We apply the criteria C 1 , C 2 <strong>and</strong> C 3 to a subgraph<br />
t 1 (→x 1 )t 2 (→x 2 ) t 3 (→l 2 )a 1 associated with agent A. This<br />
subgraph represents a task t 1 exp<strong>and</strong>ed to a task t 2 which<br />
is in turn exp<strong>and</strong>ed to a task t 3 <strong>and</strong> allocated to an asset<br />
a 1 . Here, we assume that t 1 has a parent external to A. We<br />
further assume that the asset a 1 is an organizational asset<br />
representing agent B. The task node t 1 satisfies C 1 <strong>and</strong><br />
hence is aggregated into an abstract node n 1 . Similarly,<br />
the task node t 3 satisfies C 3 <strong>and</strong> is aggregated into a node<br />
n 2 . The subgraph (→x 2 )t 2 (→x 2 ) therefore matches C 2 <strong>and</strong><br />
is associated with a single edge e 1 ∈E’ which connects<br />
the two abstract nodes n 1, , n 2, ∈ N’.<br />
Together, the equivalence criteria for abstract nodes <strong>and</strong><br />
edges leads to the identification of equivalent subgraphs.<br />
By changing the equivalence criteria for aggregated nodes<br />
<strong>and</strong> edges, a variety of different graph reduction<br />
mappings can be achieved.<br />
The computational complexity of the approach described<br />
above varies depending on the complexity of finding <strong>and</strong><br />
matching isomorphic subgraphs. Cougaar task graphs are<br />
generally well structured, <strong>and</strong> for most of the equivalence<br />
criteria that are described in the following, subgraphs can<br />
be matched using a (worst case) O(n) graph traversal,<br />
where n is the size of the subgraph. In the equivalence<br />
criteria described in the next section, all subgraphs fall<br />
within a single agent’s plan graph, thus bounding the size<br />
of the matched subgraphs. If m is the total number of<br />
abstract nodes discovered, then the total computational<br />
complexity is O(m * n).<br />
3.1 Algorithm implementation <strong>and</strong> applications<br />
The implementation of the graph reduction algorithm<br />
within Castellan allows generation of the task graph from<br />
an arbitrary stream of events. Except for asset <strong>and</strong><br />
organizational information, it is not necessary to have a<br />
complete plan graph to use this approach to devise<br />
reduced graphs.<br />
The following types of reduced graphs were found to be<br />
useful for underst<strong>and</strong>ing Cougaar societies <strong>and</strong> are<br />
implemented within the Castellan system.<br />
Aggregate task graphs defined using an equivalence<br />
criterion that maps tasks of the same type with the<br />
identical verb to single nodes. Also, theis equivalence<br />
criterion requires a strict ordering in depth between tasks<br />
which are aggregated. Specifically, in order to map a<br />
node n to an abstract node n’, the node n’s parents (<strong>and</strong><br />
all of its ancestors by implication) must map to an<br />
abstract node n 2 ’ which is an ancestor of n’. This<br />
requirement is imposed to prevent cycles from appearing.<br />
Figure 2 shows a conceptual representation of task<br />
aggregation in which multiple “similar” subgraphs are<br />
collapsed into a single aggregate subgraph. An example<br />
of a task aggregate plan graph is shown in Figure 3.<br />
Asset dependency graphs consider the assets (both<br />
organizational <strong>and</strong> physical) as abstract nodes. In this<br />
case, no aggregation of assets is performed as all assets<br />
are considered unique. The criteria for abstract edges are<br />
as follows:<br />
• All assets are mapped to abstract nodes.<br />
(Optionally, additional asset matching criteria<br />
can be introduced to aggregate assets.)<br />
• In addition, all agents that generate tasks are<br />
designed as “Source” abstract nodes. (These<br />
serve as the roots of the reduced DAG.)<br />
• All tasks <strong>and</strong> allocations that form plan graph<br />
dependencies between different assets are<br />
mapped to a single abstract edge.<br />
Asset dependency graphs are useful for finding both<br />
organization <strong>and</strong> physical dependencies within a<br />
distributed plan.<br />
Workflow graphs characterize the input/output<br />
relationships between agents <strong>and</strong> are particularly useful