Page 155 -

144 M. Adams
level interaction. The system also supports forward and backward “jumps” through
a process instance, but only by authorized staff who instigate the skips manually
[204].
The AdaptFlow prototype [111] provides a hybrid approach to flexibility. It sup-
ports the dynamic adaptation of process instances, although each adaptation must
be confirmed manually by an authorized user before it is applied (alternate manual
handling to override the dynamic adaptation offered is also supported). Also, the
rule classifications and available exception handling actions are limited to medical
treatment scenarios. The prototype has been designed as an overlay to the ADEPT
system, providing dynamic extensions.
The ADOME system [57] provides templates that can be used to build a workflow
model, and provides some support for (manual) dynamic change; it uses a central-
ized control and coordination execution model to initiate problem solving agents to
carry out assigned tasks. A catalog of “skeleton” patterns that can be instantiated or
specialized at design time is supported by the WERDE system [54]. Again, there is
no scope for specialization changes to be made at runtime.
AgentWork [176] provides the ability to modify process instances by dropping
and adding individual tasks based on events and ECA rules. However, the rules do
not offer the flexibility or extensibility of the YAWL approach, and changes are lim-
ited to individual tasks, rather than the task-process-specification hierarchy. Also,
the possibility exists for conflicting rules to generate incompatible actions, which
requires manual intervention and resolution.
The ActivityFlow specification language described in [151] divides workflows
into different types at design time (including ad-hoc, administrative, or produc-
tion), and provides an open architecture that supports interaction and collaboration
of different workflow systems. The system, like ADEPT, advocates the use of a
dedicated (human) workflow coordinator/administrator to monitor workflows with
an eye on deadlines, handle exceptions, prioritize, stop, resume and abort processes,
dynamically restructure running processes, or change a specification.
An approach that uses a “society of intelligent agents” that work together to exe-
cute flexible processes is found in [257], and another that uses BPBots (Business
Process Robots) to perform the roles of service requesters, providers, and brokers in
the formation of a hierarchical community for the execution of a process instance is
introduced in [279]. A further approach using incompletely specified process defini-
tions is found in the SwinDeW (Swinburne Decentralized Workflow) project [278].
SwinDew is a peer-to-peer based decentralized model, where a process definition
is split into a set of task partitions and distributed to peers, and on-the-fly pro-
cess elaboration is performed at runtime. Thus, a multi-tiered process modeling and
execution framework is provided.
CBRFlow [258] uses a case-based reasoning approach to support adaptation of
predefined workflow models to changing circumstances by allowing (manual) anno-
tation of business rules during runtime via incremental evaluation by the user. Users
must be actively involved in the inference process during each case. An approach,
which integrates CBRFlow into the ADEPT framework, is described in [213]. In
doing so, semantic information about the reasons for change, and traceability data,