Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
www.elsevier.com/locate/knosys
Abstract
This paper argues that AI planning is a technology ripe for use on real world problems as shown by a number of current applications. An
introduction is given to AI planning, and then followed by discussion of the successful application of this technology to generating operating
procedures for chemical plants. A description is given of the methodology for developing planning domains and finally the results of its
application to operating procedure synthesis are discussed. 䉷 2000 Elsevier Science B.V. All rights reserved.
Keywords: Artificial intelligence planning; Chemical plant operating procedures; Knowledge engineering
system that generates plans from first principles using a not be possible to move B onto C to solve the other goal.
temporal domain model, a Smart Executive for executing Resolving conflicts can be carried out by reordering the
the plans, and a model-based Mode Identification and conflicting actions, inserting new actions, or by replanning,
Recovery system which carries out fault diagnosis and as discussed in Ref. [10].
suggests mode reconfigurations to the Smart Executive. During planning, the search space can become enormous
Experiments carried out during May have shown that the if no techniques are used to limit its size. Least-commitment
planning system is capable of successfully controlling the planning [11] is an approach to reducing search spaces. It
spacecraft given very high-level commands from the encompasses non-linear planning, in which only essential
ground, and of coping with both equipment and sensor ordering constraints between actions are introduced, leaving
failures. all others unordered (in pseudo-parallel), allowing a whole
The domains of these systems are all rather different: set of plans to be represented at once. It also includes
what they have in common is that they solve difficult and constraining the possible instantiations of an object used
potentially combinatorial sequencing problems. We will in the plan rather than committing to a particular instantia-
show that generating operating procedures for chemical tion. Hierarchical Task Network (HTN) planning also
process plant has these same characteristics, and we argue reduces the search space by representing a problem as a
that AI planning is a good solution for problems involving hierarchy of tasks that need to be achieved, allowing a
difficult and potentially combinatorial sequencing problems. plan or part of a plan to be represented by a high level of
This success forms part of the impetus behind an increas- abstraction, with the lower levels, and more detailed part of
ing number of international planning-centred events: the the plan, left for later expansion.
biennial international AI Planning and Scheduling Confer- In this paper, we will look at how AI planning has been
ence, the transformation of the European Planning and successfully applied to a real word problem of generating
Scheduling Workshop series into a full conference from operating procedures for chemical plants.
1998 and the recent (October 1998) setting up of an EU
funded network of excellence for planning. PLANET
(http://planet.dfki.de/).
3. Operating procedure synthesis
therefore not only desirable for the commissioning team, but through its post-conditions—and should therefore always
as a tool to explore operability during plant design. be the same. Valve operations violate this assumption [20].
Therefore, valve sequencing is handled by a specialist
module in CEP that uses an approach that we call ‘action
3.1. Chemical engineering planner synergy’ and is based on work by [14,21]. A maze searching
algorithm is used to find a route for a flow between given
The Chemical Engineering Planner (CEP), has been
start and end points. All the valves around this route are then
developed over the last five years, initially as a PhD project
closed and finally those actually on the route are opened.
[18] and in the last three years as part of the EPSRC-funded
Thus, CEP can be seen as a general-purpose AI planner with
INTergrating OPerability (INT-OP) project. 1,2 CEP has
domain-related specialist additions.
been developed incrementally through case studies of
increasing scope and complexity, and as we show is now
more capable than any of the systems referenced earlier. We
3.2. OPS domains
will only summarise the structure of CEP in this paper.
CEP divides the tasks involved in OPS into three areas:
An incremental approach has been taken to our case
planning using operators, the handling of safety considera-
studies. Initially we looked at ‘toy’ problems discussed in
tions and valve sequencing. The first two of these three areas
OPS literature. The next level of case study was a test rig in
are handled by a state-of-the-art least-commitment planner
the Chemical Engineering department at Loughborough
[11], which uses the concept of ‘goals of prevention’ [19] to
University, a Double Effect Evaporator (DEE).
prevent actions being incorporated into the operating proce-
Finally, the successful completion of the DEE case study
dure that will take a plant through any unsafe states. Safety
led to another move up in domain complexity to actual
is clearly a particular concern in a chemical plant domain: a
chemical plant case studies, an ICI ammonia plant and a
plan which moves the plant to a desired end-state is unac-
BP acetic acid plant.
ceptable if—for example—along the way explosive gases
In this paper, we will describe the results from the DEE
have been mixed together. ‘Goals of prevention’ are defined
with the chemical plant case studies discussed in less detail
as safety restrictions as part of the overall description of the
due to commercial confidentiality. A brief outline of the
plant.
process for the DEE test rig is shown in Fig. 1. In brief, it
CEP deals with valve sequencing as a special case. A
uses evaporation in two successive evaporating vessels to
characteristic of the opening and closing of valves in a
extract salt from saline solution. Steam from the first
chemical plant—actions required in order to produce
evaporation process is used as a heat input to the second.
flows of chemicals to specified vessels or other compo-
Because this was a teaching rig, at the end the salt is
nents—is that the effect of the action at a particular valve
dissolved back into the water so that the process can run
is dependent on associated actions at other valves. However,
indefinitely.
an assumption of the standard AI planner representation of
The DEE is a complex domain and is much nearer to a
actions is that the effect of an action should be represented
real-world chemical plant than the domains used in previous
1
Academic Partners: Loughborough University and Salford University.
work. Not only does the DEE set-up contain a larger number
2
Industrial Partners: BG Plc, BP, Cogsys Ltd, ICI, Science Systems Ltd, of components (67) than in most previous domains, but the
and TCCL. number of different types of equipment is also large (15),
64 R.S. Aylett et al. / Knowledge-Based Systems 13 (2000) 61–69
with valves, controllers, pumps, heaters, coolers, evapora- developed during earlier work at Loughborough [23] to
tors, feed tank, mixing tank and a barometric condenser. model individual components in a particular plant. We
soon found that the manual entry of these descriptions for
every component in a particular plant using CEPs syntax is
4. Planning methodology
non-trivial: it is both very time-consuming and prone to
error. An automatic system was therefore developed for
Two closely coupled steps are involved in applying a
producing the domain description. A popular drawing pack-
planner to a new domain: knowledge acquisition and
age, AutoCAD, has been adapted to provide the standard
domain modelling. In the DEE case study, knowledge was
chemical engineering equipment symbols for the user to add
acquired by reading the documentation on the test rig, visit-
the name and connections for it. Thus on completion of the
ing the DEE installation, and by interviewing a Loughbor-
drawing, the necessary information has been collected to
ough University colleague with an understanding of the
allow the automatic creation of a domain description file.
working of the test rig. While there are important issues
here, we will not touch on them in this paper.
We will however discuss domain modelling in more 4.2. Planning operators
detail, as the amount of time and effort required to construct
The next stage is to develop the planning operators that
a particular domain model is a major obstacle to the use of
describe the actions that can be used to produce a plan.
AI planning in the solution of real-world problems [22]. If
Where the domain description gives the static content and
CEP is to be used by design engineers for real industrial
layout of a plant, planning operators define its behaviour. A
plant, it must be straightforward to construct the domain
CEP operator consists of a goal(s) that can be achieved
model. We therefore report the lessons learned from our
when the precondition(s) for the operator are true—essen-
case studies.
tially the STRIPS [24] representation still widely used in AI
After knowledge acquisition, the information acquired
planning systems in spite of all the other changes in the field
must be transformed into a form that the planner CEP can
since then. An example operator is shown in Fig. 2. This is a
understand and use. All planners require the first four parts
macro operator, which allows information about the order in
listed below in their domain model, and CEP introduces a
which the pre-conditions must be satisfied to be entered
new element called pairs:
[20]. CEP also offers directly executable primitive opera-
• domain description (plant model); tors—such as Open Valve operator—and more abstract
• planning operators (actions that can be carried out); expandible operators used for hierarchical planning [27].
• safety restrictions (constraints on states); Operator development is a time-consuming and difficult
• domain problem (the plant operations to be carried out); part of domain development and one on which there is
• pairs (variable associations). minimal guidance in the available literature. Yet the correct-
ness and efficiency of the planning process in a domain
4.1. Domain description depends very heavily on operator definition.
A planning operator can be more or less generic in nature.
A planner will require a description of what is in the For example, in a domain for moving blocks around, a move
domain—in this case a chemical process plant. CEP uses operator could apply only to block4, to all red blocks, to all
an implementation of a hierarchical frame-based description blocks of a given size, to any regular polygonal piece of
R.S. Aylett et al. / Knowledge-Based Systems 13 (2000) 61–69 65
DEE domain. A single model of the domain allowed proce- tions can be set up to force CEP to use the Glass Preheater
dures to be created for the start-up, shutdown and the isola- this is a rather counter-intuitive use of this feature.
tion of pieces of equipment for maintenance.
The start-up procedure generated by CEP for the DEE 5.3. Linking and ordering actions
contained 52 steps, with 19 operators used in planning.
The time taken to generate the operating procedure was The output of a partial-order planner such as CEP is a
under 5 s on a Sparc 5. An example section of a operating plan-network in which only those actions, which must
procedure generated by CEP can be seen in Appendix A. follow each other are ordered with respect to each other.
Moreover CEP proved capable of finding alternative proce- Actions, which may be taken in any order, appear unordered
dures via backtracking at the user’s request. in parallel branches of the plan-net. Thus it is easy to see the
critical path in this representation, and one of the case-study
5.1. Comparison with previous systems plants used hand-produced networks to help operators
understand this critical path.
The planner CEP has been used to produce operating However, it is usual for the last step in an AI planner to be
procedures using AI planning for a domain more complex the transformation of the plan-net into a linear sequence,
than any other previously attempted. The work of the early since usually only one-step-at-a-time execution is planned
1980s [13,28] using state-graphs limited sample problems to for. This process is known as linearisation. The plan-net,
plants containing a handful of valves because of the number supports user interaction in the linearisation process since
of states they generated: 20 valves each with two states it shows what actions can be moved in a particular linear-
produces 1,048,576 nodes in a state-graph. This demon- isation and which cannot. The re-ordering of actions may be
strates that OPS is indeed a difficult and combinatorial desirable for two reasons. Firstly, grouping actions together
sequencing problem. for operating a certain piece of equipment makes the plan
Other workers used larger plant [12] but only considered clearer. Secondly, a ‘better’ plan may be produced by taking
valves and not vessels. A real-world nuclear fuel processing the actions in a certain order. For example, if two valves
plant was used in [17], but this work concentrated on opti- have to be opened manually, then these actions should be
mising a hand-generated plan. CEP has successfully solved together if they are geographically next to each other in the
every sample problem reported in the OPS literature except plant. CEP cannot currently take geographical proximity
for those requiring numerical calculations. We therefore into account, since the ELD from which it works contains
argue that CEPs success in our case studies demonstrates only the topological relationship of equipment.
a big step in the state-of-the-art for OPS. However CEPs linearisation routine does group
sequences of actions together which contribute to the
5.2. Quality of results same high level goal. This results in the output appearing
in chunks related to the operation of certain parts of the plant
The procedures produced were evaluated by the domain and allows the human operator to understand it more easily.
experts who had been used for the knowledge acquisition. Indentation is used in the output to show how primitive
The expert for the DEE found the generated procedures actions in the procedure contribute to overall goals. An
adequate—in the sense that the start-up procedure would extension would allow the user to reorder actions manually,
successfully start up the plant. This is an important result for subject to the constraints of the plan net, seen as highly
a domain of this complexity and validates the overall desirable by prospective users.
approach of using AI planning technology on this problem.
However, he also found the generated procedures in some 5.4. Generic operators
ways naive—in the sense that they were not always iden-
tical to the ones an expert would produce. In the DEE case study, every component is now operated
A major example of this concerned the use of a glass by a generic operator, all of which are found in a component
preheater in the DEE plant. It is possible to start up the operator library. Moreover, thirteen of these operators were
plant without using this preheater, and accordingly CEP reused for a subsequent case-study using the back-end loop
originally generated a procedure that did not use it. The section of the ICI ammonia plant, and six for the corrosion
reasons for using the preheater during start-up are: the metal removal system of the BP acetic acid plant, demon-
temperature of the brine can be increased in stages, protect- strating the value of a library of planning operators.
ing the glass lined vessels, and the control of the tempera- A more important issue concerns the extent to which real-
ture of the brine entering the first evaporator is easier with world components in process plant are generic. For exam-
two heaters. An expert in operability, seeing that the design ple, of the two heat exchangers in the DEE, one has an extra
contained a glass preheater, would infer that it was there for port for output of the steam used to heat the material passing
the purpose of start-up and accordingly use it. This oper- through the exchanger. In real plants, components are
ability knowledge is not easy to represent within the sourced from a variety of manufacturers and so there may
confines of planning operators, and though safety restric- be differences in the way each is constructed and operated.
R.S. Aylett et al. / Knowledge-Based Systems 13 (2000) 61–69 67
If this variation is empirically shown to be very large, a planning to solve real-world problems. Planning attacks
generic library of operators might be impossible. problems more complex than the average KBS, and there-
An obvious approach to this problem is to consider fore tools are much more important. The automatic genera-
attaching the library of operators more firmly to the compo- tion of the domain description for CEP from AutoCAD was
nent hierarchy, with the use of object-oriented inheritance one tool developed, and the library of generic operators
and specialisation mechanisms to control variation. A front- derived from the case study is another even more important
end interface has therefore been developed to allow the one. Not discussed here were other tools to help the user in
equipment hierarchy with its associated planning operators running CEP, collectively called CEP-Run. Finally, CEP-
to be browsed and displayed. This interface, which forms Tool was created to give further user support to the devel-
part of an overall front end called ‘CEP-Tool’, also allows opment of operators, providing operator templates, an
planning operators to be modified, or, if necessary created, operator editor and an interface for interacting with the
without any knowledge of CEPs underlying syntax. operator library. We are part of a project starting shortly
(late 1999–early 2000) called PLANFORM, funded by
EPSRC in the UK, which will be investigating this area
6. Conclusions further.
The production of library of generic operators in the DEE
A number of conclusions can be drawn from our research, case study illuminated a particular problem. It is often possi-
some specific to the domain of Process Plant Operating ble to solve problems in a particular planning domain by ad
Procedure Synthesis, and others of more general relevance hoc fixes, frequently in the planning operators. As CEPs
to industrial applications of AI planning. capabilities were increased, so it became possible to solve
A positive conclusion is that the case studies validate the each problem in a more general and principled way. Thus
use of state-of-the-art AI planning techniques in OPS. As the ability to produce a library of generic operators is not
discussed above, this has made it possible to deal success- only an indispensable tool for domain development in the
fully with a large and complex domain where conventional future, it is also, we argue, indirectly a measure of the
KBS technology might well have struggled with the combi- adequacy of a planner.
natorial sequencing. Note that this was achieved without We argue that AI planning technology has now reached a
incorporating explicit reasoning about time or resources level of maturity where it can be successfully applied to
but by sequencing alone. We believe that in a real applica- difficult real-world problems. Just as KBS technology in
tion, a back-end scheduling system along the lines of [17] general has made a powerful contribution to the manage-
could be used to look at optimisation of procedures and even ment of manufacturing systems, so AI planning has the
the generation of control system specifications for automatic potential to solve problems in this area previously seen as
plant. Planning operators sometimes contain some implicit too complex to be tackled successfully.
qualitative reasoning—for example the need to wait until a In particular, OPS is a new applications area for AI plan-
fill reaches a given level, but OPS is not process design, ning, but we suggest one of considerable promise. The INT-
where quantitative methods would be inescapable. OP project developed a system that can be used in a real-
During the case studies, it became clear that number of industrial environment to produce operating procedures and
improvements were needed: in the linearisation process, in with further development could be used to produce operat-
the valve sequencing component and in the incorporation of ing procedures earlier in the plant life-cycle with real
general operability knowledge: all of these improvements savings in time and effort. Work by other groups on the
have now been made. We argue that this indicates that related areas of production planning [30] and generating
though general-purpose hierarchical least-commitment plant control sequences [31] indicates that this technology
planning algorithms are powerful, real domains also require has many potential uses in manufacturing. We also note that
domain-specific problem-solving along with a great deal of plant operating procedures are only one example of the need
domain-specific knowledge. Hopefully, in the same way as for accurate, efficient and safe procedures in industry and
work in knowledge-engineering methodologies has identi- commerce and see many possibilities for extending the AI
fied specific approaches to different types of diagnosis, work planning approach to the generation of other types of opera-
in a wider range of real-word planning domains will begin to tional procedure.
establish abstract categories in such domains which will
support a more principled approach to the choice of plan-
ning technologies for particular problems [26,29]. Appendix A
A number of knowledge engineering issues arose from
the DEE case study. Firstly, the time and effort required to 1. Set controller LRC_5 and turn on.(46)
develop the original domain was substantial, with about 20 2. Set controller FRC_6 and turn on.(45)
person-days of effort involved in developing the operators 3. Set controller LRC_7 and turn on.(44)
without using any tools. Development of tools to assist in 4. Set controller FRC_8 and turn on.(43)
domain development is, we believe, vital to the use of AI 5. Set controller TRC_9 and turn on.(42)
68 R.S. Aylett et al. / Knowledge-Based Systems 13 (2000) 61–69
[20] R.S. Aylett, J. Soutter, G.J. Petley, P.W.H. Chung, A. Rushton, AI [26] R.S. Aylett, S.D. Jones, Planner and domain: domain configuration
planning in a chemical plant domain, Proceedings, ECAI’98, 1998, for a task planner, International Journal of Expert Systems 9 (2)
pp. 622–626. (1996) 279–318.
[21] E. O’Shima, Safety supervision of valve operations, Journal of [27] R.R. Aylett, G.J. Petley, P.W.H. Chung, J. Soutter, A. Rushton, Plan-
Chemical Engineering of Japan 11 (5) (1978) 390–395. ning and chemical plant operating synthesis: a case study, in: S. Steel,
[22] S.A. Chien, R.W. Hill, X. Wang, T. Estlin, K.V. Fayyad, H.B. R. Alami (Eds.), Recent Advances in AI Planning, Springer Lecture
Mortenson, Why real-world planning is difficult: a tale of two appli- Notes in AI, vol. 1348, 1997, pp. 39–51.
cations, in: M. Ghallab, A. Milani (Eds.), New Directions in AI Plan- [28] A. Kinoshita, T. Umeda, E. O’Shima, An algorithm for synthesis of
ning, IOS Press, Washington, DC, 1996, pp. 287–298. operational sequences of chemical plants, 14th Symposium on
[23] P.W.H. Chung, Qualitative analysis of process plant behavior, in: Computerized Control and Operation of Chemical Plants, Vienna,
P.W.H. Chung, G. Lovegrove, A. Ali (Eds.), Proceeding Industrial Austria, 1981.
and Engineering Applications of AI and Expert Systems, Gordon and [29] A. Valente, Knowledge-level analysis of planning systems, ACM
Breach, London, 1993, pp. 277–283. SIGART Bulletin Special Issue 6 (1) (1995) 27–61.
[24] R.E. Fikes, N.J. Nilsson, Strips: a new approach to the application of [30] D.E. Wilkins, Can AI planners solve practical problems? Computa-
theorem-proving to problem-solving, Artificial Intelligence 2 (1971) tional Intelligence 6 (4) (1990) 232–246.
189–208. [31] L. Castillo, J. Fedz-Livares, A. Ganzalez, An AI-based tool for the
[25] G.J. Petley, R.S. Aylett, P.W.H. Chung, A. Rushton. Development of automatic synthesis of control programs, Second International
a reusable operator library for chemical plant domains proceedings, Symposium of Manufacturing Systems, 1998.
17th Workshop of the UK Planning and Scheduling SIG, University
of Huddersfield, ISSN. 1368-5708, 1998, pp. 145–156.