132 t wagner and v lesser returning to the example

132 T. Wagner and V. Lesser
Returning to the example, Get-Basic has two actions, joined under the sum() quality-
accumulation-function( QAF ), which de fi nes how performing the subtasks relate to per-
forming the parent task. In this case, either action or both may be employed to achieve
Get-Basic . The same is true for Gather-Reviews . The QAF for Build-PC-Product-Objects
is a seq last() which indicates that the two subtasks must be performed, in order, and
that the quality of Build-PC-Product-Objects is determined by the resultant quality
of Gather-Reviews . There are nine alternative ways to achieve the top-level goal in
this particular sub-structure.
4
In general, a T Æ MS task structure represents a family
of plans, rather than a single plan, where the different paths through the network ex-
hibit different statistical characteristics or trade-offs. The process of deciding which
tasks/actions to perform is thus an optimization problem rather than a satisfaction prob-
lem.
T Æ MS also supports modeling of tasks that arrive at particular points in time, par-
allelism, individual deadlines on tasks, earliest start times for tasks, and non-local tasks
(those belonging to other agents). In the development of T Æ MS there has been a con-
stant tension between representational power and the combinatorics inherent in working
with the structure. The result is a model that is non-trivial to process, coordinate, and
schedule in any optimal sense (in the general case), but also one that lends itself to
fl exible and approximate processing strategies. This element of choice and fl exibility is
leveraged both in designing resource-bounded schedules for agents and in performing
online scheduling in a resource bounded fashion.
QUALITY
10
20
30
50 100
Histogram of Quality - Action Mx
100% Density
FINISH_TIME
10
20
30
40
5 10 20 15
Histogram of Finish_Time - Action Mx
Exceeds
Deadline
10% of the Time
Quality
10
20
30
50 100
Histogram of Modified Quality
Action Mx
90% Density
10% Density
0
Fig. 3. Re fl ecting Probability of Missing Deadline in Method Quality
3 Modeling and Reasoning about Temporal and Resource
Constraints
T Æ MS tasks may have both soft and hard constraints that must be considered when
scheduling. In terms of hard temporal constraints, any T Æ MS task may have a hard
deadline, by which some quality must be produced (or it is considered a failure), as well
4
While it might appear per the seq last() QAF that there are only two possible resultant quality
distributions, the enables interaction between Build and Gather affects the possible quality
values for Gather .

Design-to-Criteria Scheduling: Real-Time Agent Control 133
as an earliest-start-time, before which the task may not be performed (or zero quality
will result). These hard constraints may also be caused by hard commitments
5
made
with other agents or hard delays between task interactions. The constraints may also be
inherited from nodes higher in the structure – thus a client may specify a hard deadline
on the Build-PC task that applies to all subtasks, or a deadline may be speci fi ed on the
process of Gathering-Reviews . If multiple temporal constraints are present, the tightest
or most conservative interpretation applies.