MANAGING E-MARKET TRANSACTION PROCESSES
Exploring the limits of process management with a multiagent system
John Debenham
University of Technology, Sydney, Australia
Keywords: Multiagent systems, process management, datam
ining, electronic market transaction management
Abstract: Knowledge-driven processes are business proces
ses whose execution is determined by the prior knowledge
of the agents involved and by the knowledge that emerges during a process instance. They are characteristic
of emergent business processes. The amount of process knowledge that is relevant to a knowledge-driven
process can be enormous and may include common sense knowledge. If a process’ knowledge can not be
represented feasibly then that process can not be managed; although its execution may be partially
supported. In an e-market domain, the majority of transactions, including requests for advice and
information, are knowledge-driven processes for which the knowledge base is the Internet, andso
representing the knowledge is not an issue. These processes are managed by a multiagent system that
manages the extraction of knowledge from this base using a suite of data mining bots.
1 INTRODUCTION
In an experimental e-market, transactions
(Debenham, 2001) include: trading orders to buy
and sell in an e-exchange, single-issue and multi-
issue negotiations between two parties, requests for
information extracted from market data as well as
from news feeds and other Internet data. This
e-market is used at UTS for research and teaching.
In it every market transaction is managed as a
business process (Fisher, 2000). To achieve this,
suitable process management machinery has been
developed. To investigate what is ‘suitable’the
essential features of these transactions are related to
two classes of process that are at the ‘high end’ of
process management feasibility (van der Aalst et al.,
2001). The two classes are goal-driven processes
(Sec. 2) and knowledge-driven processes (Sec. 3).
The term “business process management” is
generally used to refer to the simpler class of
workflow processes (Fisher, 2000), although there
notable exceptions using multiagent systems
(Jennings et al., 2000). Sec. 4 describes the
relationship between the transactions themselves and
the contextual information extracted from the
Internet and from market data. Sec. 5 discusses the
single-issue and multi-issue negotiation transactions.
e-market transactions are described in Sec. 6.
2 GOAL-DRIVEN PROCESSES
A goal-driven process has a process goal, and
achievement of that goal signals the termination of
the process. The process goal may have various
decompositions into possibly conditional sequences
of sub-goals where these sub-goals are associated
with (atomic) activities and so with atomic tasks.
Some of these sequences of tasks may work better
than others, and there may be no way of knowing
which is which (van der Aalst et al., 2001). A task
for an activity may fail outright, or may fail to
achieve its goal in time. In other words, a central
issue in managing goal-driven processes is the
management of task failure. Hybrid multiagent
architectures whose deliberative reasoning
mechanism is based on “succeed/fail/abort plans”
(Rao et al., 1995) are well suited to the management
of goal-driven processes. Goal-driven processes are
a more powerful concept than production workflows
(or, called “activity-driven processes” in
(Debenham, 2000)). Activity driven-processes are
associated with possibly conditional sequences of
activities where performing that sequence is
assumed to “work always”.
Following (Rao et al., 1995) a plan for a goal-
di
rected process can not necessarily be relied upon
to achieve its goal even if all of the sub-goals on the
322
Debenham J. (2004).
MANAGING E-MARKET TRANSACTION PROCESSES - Exploring the limits of process management with a multiagent system.
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 322-330
DOI: 10.5220/0002597203220330
Copyright
c
 SciTePress

chosen path through that plan have been achieved.
The success condition (SC), described in
(Debenham, 2000), is a procedure whose goal is to
determine whether a plan’s goal has been achieved.
The final sub-goal on every path through a plan is
the plan’s success condition. The success condition
is a procedure; the execution of that procedure may
succeed(3), fail (7) or abort (A). If the execution of
the success condition fails then the overall success
of the plan is unknown (?). So the four possible
plan exits resulting from an attempt to execute such
a plan are as shown in Fig. 1. A plan body is
represented as a directed AND/OR graph, or state-
transition diagram, in which some of the nodes are
labelled with sub-goals.
SC


if(G)
if(~G)

?

p
lan for goal G
A
A
A
Figure 1: The four plan exits
Process Goal
(what we are trying
to achieve over all)
Next-Goal
(what to try to
achieve next)
Initialise
?not SC and not
activity goal?
Select
Identify
Back-up
Identify
?SC?
?activity
goal?
Activity
Do it
Select
Evaluate it
Plan
Figure 2: A simplified view of goal-driven process
management
The management of goal-driven processes is
shown in a simplified form in Fig. 2. There, starting
with the overall process goal, repeated
decomposition of plans and goals is performed until
either the next goal is a success condition or is an
activity goal—ie: a goal for which there is a hard-
coded procedure. Fig. 2 is simplified because it does
not show what happens if the success condition
returns fail “7”, or what happens if a plan is aborted.
Further it does not show the mechanism for selecting
plans for goals. For goal-driven processes there is,
in general, no ex ante ‘best’ choice of plan.
3 KNOWLEDGE-DRIVEN
PROCESSES
A second class of process, whose management has
received little attention, is called knowledge-driven
processes. Process knowledge is all the knowledge
that is relevant to a process instance. It includes
common-sense knowledge, knowledge that was
available when an instance is created, and
knowledge acquired during the time that that
instance exists. A knowledge-driven process may
have a process goal, but the goal may be vague and
may mutate. In so far as the process goal gives
direction to goal-driven—and activity-driven—
processes, the process knowledge gives direction to
knowledge-driven processes. The body of process
knowledge is typically large and continually
growing—for example, it may include common
sense knowledge—and so knowledge driven
processes are seldom considered as candidates for
process management. They are typically supported,
rather than managed, by CSCW systems. But, even
complex knowledge-driven processes are “not all
bad”—they typically have goal-driven sub-processes
that may be handled as described above.
Knowledge-base processes are a special type of
knowledge-driven process for which the process
knowledge can be represented and accessed by a
process management system. This proves to be a
useful concept for managing e-market transactions.
Process Goal
(the current over
all goal)
Process Knowled
g
e
(knowledge of all that
is relevant to the
process instance)
Next-Goal
(what to try to
achieve next)
Activity
(what should happen next)
Decompose
(in the context of the
process knowledge)
Do it —
(until termination
condition satisfied)
New Process
Knowledge
Add to
Revise
Select
Initial process Goal
Initialise
Figure 3: A simplified view of knowledge-driven process
management
The management of goal-driven and knowledge
driven processes are radically different. Goal-driven
processes may be managed by a goal/plan
decomposition process (see Fig. 2), and knowledge-
driven processes are managed by continually
reviewing the growing corpus of process
knowledge—this is illustrated in Fig. 3. That Figure
is deceptively simple in that the business of
managing the process knowledge and of revising the
process-goal andnext-goal in the light of that
growing body of knowledge is far from trivial in
MANAGING E-MARKET TRANSACTION PROCESSES: Exploring the limits of process management with a multiagent
system
323

even simple examples. In general this problem will
be intractable. But in some cases, including the
majority of e-market transactions, smart tools may
be used to do this. This is discussed in the next
section.
4 E-MARKET TRANSACTIONS
AND CONTEXTUAL
INFORMATION
E-market transactions include: trading orders to buy
and sell in an e-exchange (single-issue and multi-
issue negotiations as described above), requests for
market data as well as requests for information
extracted from news feeds and other Internet data.
In an experimental e-market, all e-market
transactions are managed as constrained knowledge-
driven processes.
Sec. 5.1 discusses single-issue one-to-one
negotiation. Single-issue negotiation also takes
place in exchanges, for example a ‘buy’ trading
order to “buy a chair and a desk for less than $100”.
This is represented (see Fig. 4) as a naive plan with
goal [G, c] = [desk and chair have been purchased,
cost < 100]. This plan has sub-goal SG
1
= ‘chair
and desk selected’, [SG
2
, c
2
] = [chair purchased,
cost < 30], [SG
3
, c
3
] = [desk purchased, cost < 50],
and [SG
4
, c
4
] = [desk and chair delivered, cost <
20]. The management of this purchase order is
represented as a plan whose goals have monetary
constraints.
An example of a request for information is “find
out all you can about ABC Corp within five
minutes”. This triggers a process to locate, extract,
validate, condense and combine information from
the Internet. The location and extraction tasks are
achieved by data/text mining bots that are described
in [7]. A handcrafted plausible inference network
combines contradictory information. The use of
belief nets that can be trained “off line” is very
tempting and is currently being investigated [8].
The data/text mining bots produce output in the form
[ data, belief ]—ie: some data and a measure of the
belief held in the validity of that data. A request for
information is first represented as a goal/constraint
pair: [ find_info_about(‘ABC Corp’):
time_upper_limit = now + 5mins ]. Given a
goal/constraint pair, a plan (see Fig. 4) is selected
for it—the mechanism for selecting a plan is
described in Sec. 6. A plan for a goal/constraint
pair is a possibly conditional state-chart of sub-goals
over which constraints are distributed as described in
Sec. 6. For a ‘find_info_about’process, the plan
uses a Dempster-Shafer network (see Fig. 5) to
combine results [D
i
, b
i
], in the form [ data, belief ],
extracted from the Internet by a suite of data/text
mining bots. The network actually does more than
combine information. If the level of belief, b
R
, in a
result, R, derived by the network is below a set
threshold then a ‘reverse calculation’ identifies
‘inputs’ whose belief levels are responsible for the
low level of belief in R. Then further data/text
mining is initiated in an attempt to raise this level of
belief at least for future calculations if not for the
present calculation.
A three-year research project commencing in
2002 at UTS, is investigating the mechanisms
required to support the evolutionary process in
e-markets (Debenham, 2001). It is presently funded
by four Australian Research Council Grants;
awarded variously to the author and to Dr Simeon
Simoff:
http://www-staff.it.uts.edu.au/
~emrktest/eMarket/
Market evolution is linked to innovation and
entrepreneurship (in its technical, economic sense).
Present plans for the three year project are: (1) to
build an e-marketplace trader’s workbench that, in
principle, enables a trader to operate without
external information, (2) to assist a trader to identify
arbitrage opportunities triggered by the occurrence
of rare events, (3) to assist a trader to identify
innovative forms of trade, and, possibly, (4) to
understand somethingof the evolutionary process
itself.
5 NEGOTIATION
TRANSACTIONS
Negotiation is a process whereby two or more agents
reach an agreement on a set of issues. One-to-one
negotiation, in which there are just two negotiating
agents, is sometimes called bargaining, or
informally “haggling” or “dickering”. An issue is
any good or service that one agent can provide to
another, including money. The issue set is the
range of possible issues that may be considered
during a negotiation. An issue set may be fixed; for
example, in a single-issue negotiation where the
only issue on which agreement is sought is an
amount of money. An open issue set may contain
any issue. In alimited issue set, the issues that may
be included in an offer is limited to those chosen
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
324

from a set agreed to by the negotiating agents. An
issue—for example, “period of warranty”—is
normally associated with some value—for example,
“two years”. An offer consists of a particular set of
issues chosen from the issue set, together with
values for those issues. During a negotiation with an
open or limited issue set the collection of issues in
an offer may mutate although in practice it tends to
be moderately stable.
[G, c]
[SG
1
, c
1
]
[SG
3
, c
3
]
[SG
2
, c
2
]
[SG
4
, c
4
]
Figure 4. A plan for goal [G, c]
[D
1
, b
1]


[D
2
, b
2]





[D
n
, b
n]
[R, b
r
]
Figure 5. Belief network combines informatio n
A negotiation mechanism specifies how a
negotiation may proceed; they are sometimes called
“interaction protocols” in multiagent systems work
(Weiss, 1999). Two forms of negotiation
mechanism have received a considerable amount of
attention in the economics literature, and in
e-Markets research. First, single issue negotiation
for one or moreitems being offered either to a set of
buyers or to a set of sellers; see, for example, the
extensive work on forward and reverse auction
mechanisms (Klemperer, 2000), and second, one-to-
one negotiation, or bargaining, mechanisms.
Management of the negotiation process in an
e-Market—both for negotiation through e-exchanges
and through single- and multi-issue one-to-one
negotiation—includes a continual investigation of
the negotiation context as well as the construction,
evaluation and revision of offers. For example, the
bone fide of the opponent may require verification,
the quality of the goods should be confirmed,
alternatives should be investigated, and so on. In the
experiments described here this information is
assumed to be available on the Internet. A good
e-market negotiator should conduct these contextual
investigations as an integral part of the negotiation
process (Debenham et al., 2002). “Good
negotiators, therefore, undertake integrated
processes of knowledge acquisition combining
sources of knowledge obtained at and away from the
negotiation table. They learn in order to plan and
plan in order to learn” (Watkins et al., 2002). In the
management of the negotiation process described
here, the information and the offers develop in
tandem; they both feed off each other. The term
“e-marketplace” is used here to acknowledge this
duality between offers and contextual information.
An e-Marketplace is a market in which trading can
be conducted over the Internet, and for which
sufficient information to trade “well” is available
over the Internet. This information may be derived
from on-line market data, for mining historic market
data, from text-mining news feeds and so on. The
Sydney Stock Exchange is an example of an
e-marketplace.
Figure 4: A plan for goal [G,c]
5.1 Single-issue negotiation
Single-issue negotiation is the most common form
of negotiation, in particular where the issue is price.
The number of issues in any form of negotiation,
including single-issue, can be increased if one of the
negotiating parties offers “kick backs”. For
example, an offer of two free bottles of wine for
every dozen bought provided that you have spent
more than $500 with that merchant in the previous
twelve months. This sort of offer changes what was
initially a single-issue negotiation to a multi-issue
negotiation. In this Section it isassumed that the
negotiation is strictly single-issue and that both
parties understand the meaning of the issue. For
example, such an issue could be an amount of
money. It is argued that single-issue negotiation is
appropriately managed as a knowledge-driven
process. From a process management perspective
this is interesting because the management of “real
life”knowledge-driven processes is usually
unfeasible.
Figure 5: Belief network combines information
Two important classes of bargaining
mechanisms are alternating offers mechanisms and
single-round, “one-hit” mechanisms that may be
used when the agents have determined private
valuations in advance. For example, (Myerson et
al., 1983) shows that a one-hit“split the difference
between bid and ask” mechanism should be
preferred by both buyer and seller to any other
mechanism ex ante—that is, before their private
MANAGING E-MARKET TRANSACTION PROCESSES: Exploring the limits of process management with a multiagent
system
325

valuations are actually determined. Alternatively, an
agent’s valuations may be refined as the negotiation
proceeds—in which case an alternating offers
mechanism is which information is tabled as
appropriate—this is the focus here.
The negotiation protocol used is a time-
constrained, unbounded alternating offers protocol
(Kraus et al., 2001). In this protocol two bargaining
agents exchange offers until either one agent accepts
an offer from the other agent, one agent rejects an
offer and withdraws without penalty, or one agent
exceeds an agreed time constraint on making an
offer. So negotiation using this protocol could, in
principle, proceed indefinitely—hence the
description unbounded.
Consider a transaction to purchase something.
Suppose that this transaction can be appropriately
managed by: identifying a need, selecting a good to
satisfy that need, choosing a supplier for that good
and negotiating terms for that good from that
supplier. A sequential procedure basedon this would
not be appropriate for purchasing all classes of
goods; it could, however, be suitable for purchasing
a technical book. The appropriateness of this
“purchasing procedure” is not of concern here.
Suppose further that we wish to select the “most
appropriate” good, to choose the “best supplier” and
to negotiate “acceptable” terms. In an e-marketplace
sufficient information to trade successfully in this
sense is assumed to be available on the Internet.
The use of “software bots” to assist the buying
process by extracting contextual information from
the Internet is commonplace. For many classes of
goods, bots that do some of this work are freely
available: http://www.botspot.com/
— viz: the
sections “Shopping Bots” and “Commerce Bots”.
The entire problem considered here lies beyond the
capacity of most off-the-shelf bots that at best
recommend rather than decide. Although the use of
demographic data, collaborative filtering, clustering,
or previously expressed user preferences can deliver
reasonable performance in choosing the “most
appropriate” good for the user. With current
technology it could be reasonable to give the
authority to a bot to select, order and pay for paper
stock for photocopiers, but many would be reluctant
to permit a bot to select, order and pay for a book on
Bayesian Nets, for example.
Satisfy a need N [ Need N is satisfied ]
start
need N identified


good G selected for N
supplier S chosen for G

terms T accepted for G





d
a
t
a

&
t
e
x
t
m
i
n
i
n
g
Figure 6: A goal-driven plan to satisfy a need based on
“succeed / fail” plans
In this project the contextual information from
the Internet is first extracted by a range of data and
text mining bots mostly written by undergraduates at
UTS. Some of these bots read reviews of products
in an attempt to determine the comparative inherent
quality of a good as well as its basic attributes. This
in general leads to a collection of contradictory
evidence that is combined to give coherent advice.
The approach taken to plausible inference is
described in Sec. 4
A simple, double (ie: succeed / fail) branching
plan to manage a “purchase something” transaction
is shown in Fig. 6. That plan treats the process
sequentially in that, for example, once the good is
selected then its appropriateness is not reconsidered.
This may be appropriate when the whole transaction
can be resolved quickly, but could otherwise lead to
poor decision making. That plan relies on
information from data and text mining bots to
support the decision making in the achievement of
three of its sub-goals. [Plans for those three sub-
goals are not shown here.] Despite the vital role of
the bots, that plan manages the transaction as a goal-
driven process. The management of the same
transaction as a knowledge-driven process is
described below. This is achieved by feeding the
contextual information into the reactive “abort”
conditions in the plans.
To simplify the following discussion the
operation of the data and text mining bots is hidden
in the following predicates: INeed( N ) that means:
“I need an N”, Satisfy( N, G ) that means: “good G
is the most appropriate good that satisfies need N”.
Calculation of values to satisfy these predicates may
take some time.
Fig. 7 shows one of a sequence of linked plans
for the “purchase something”transaction. In that
Figure “d t m” denotes information that is acquired
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
326

from the Internet and databases by data and text
mining bots and combined into coherent advice by a
plausible inference network. That plan is more
intricate than the previous goal-directed version.
Even so this sequence is flawed in that the process
may now continue indefinitely; this difficulty is
addressed by constraints in Sec. 6 below. They
include reactive abort triggers that redirect the
course of the transaction if any prior decision ceases
to be valid. The direction, and possible redirection,
of this transaction is governed entirely by the
contextual information received, and repeatedly
reconfirmed, from the data and text mining bots.
This plan is useful but is not particularly noteworthy
in itself. What is of note, from a process
management perspective, is that this is a fully
managed knowledge driven process, despite its
presentation in a goal / plan framework. It is a
knowledge-base driven process where the
knowledge base is the Internet and market data, the
query mechanism is the bots, and the reactive ‘abort’
exists are used to modify the direction of the process
when necessary.
Buy G for N [ good G bought for N ]
start






good G selected for N
G bought from S
A
[ ¬INeed( N ) ]
A
[ ¬Satisfy( N, G ) ]
A
[ ¬INeed( N ) ]
d
t
m
d
t
m
Figure 7: Knowledge-driven plans to satisfy a need based
on “succeed / fail / abort” plans
5.2 Multi-issue negotiation
The discussion in Sec. 5.1 on the management of
single-issue negotiation is rather simplified in that it
assumes that the single-issue in which the offers are
expressed is unambiguously understandable. In
practice negotiation is more complicated than this.
For example, when the issue is money then an
amount expressed in dollars is readily understood,
but when and where the payment has to be made
may not be. If the issue set contains things like an
“unconditional warranty” then it would be prudent
to clarify quite what this really means. So the
feature of multi-issue negotiation that is explored
here is the opponent as a source of information, used
to clarify the meaning of an offer or otherwise.
A negotiation process with fairly minimal
functionality is shown in Fig. 8. There the process is
triggered by the arrival of an offer from the
opponent. This is analysed to ensure that the
meaning is clear—to uncover the “fine print”— and
to detect any inconsistencies. Then the offer is
evaluated to determine what it is “worth”—this can
lead to acceptance or outright rejection, or to the
development of a counter offer. Another context for
the generation of the counter offer is the history of
offers received in this negotiation—this enables an
assessment to be made of“where are the opponent is
at”. Eg: “is she about to give in?” The process
illustrated in Fig. 8 can seen as an attempt to satisfy
the high level goal “attempt to negotiate a
satisfactory outcome”. But the direction that the
process takes is determined by the flow of
information—from the offer itself, the data and text
mining bots, the opponent and from the growing
history of offers. So this is a knowledge-driven
process. At present the evaluation function is
available from the bots as described above. At the
time of writing the rest of the machinery is not yet
available, but plans are to achieve this by the end of
2002. There is much to be done, for example the
detection of inconsistencies in an offer is not trivial
even if the terms of the offer are represented in Horn
clause logic.
o
p
p
o
n
e
n
t
d
t
m
trigger
offer received
offer
analyse
offer
seek clarification
clarification
offer
evaluate
history
of offers
counter offer
accept / rejec
t
OR
develop
return
Figure 8: High-level view of the negotiation process
6 TRANSACTION CONSTRAINTS
All e-marketplace transactions are assumed to be
constrained by time constraints and possibly by cost
constraints or success constraints. Time constraints
may be the maximum (or minimum) time by which
a deal must be struck and/or by which the goods
should be delivered. The cost constraints could be
constraints on the cost of the transaction, the cost of
the goods or a combination of the two. Success
constraints may be constraints on the outcome of the
MANAGING E-MARKET TRANSACTION PROCESSES: Exploring the limits of process management with a multiagent
system
327

deal; for example, “I must have a car for the
weekend, get the best deal you can”.
The e-marketplace transaction management
system attempts manage transactions to “deliver the
best it can whilst satisfying the constraints”. To do
this it selects plans to achieve goals on the basis of
expected time and cost estimates. Further, if actual
performance differs significantly from these
estimates then estimates for subsequent plans are
adjusted leading, possibly, to a revised plan. This
will occur if network performance is unexpectedly
degraded, for example. To derive time and cost
estimates for each plan it gathers performance
measurements on each plan and sub-system, such as
an information gathering bot, and maintains running
estimates of future expected performance. It then
adjusts these estimates when measurements are
observed outside expected limits. For example, if
the network is slow when gathering data from New
York, then time estimates for extracting data from
London may be adjusted to some extent.
Time and cost performance measurements are
made for each plan and for each atomic sub-system
whenever it is used. These measurements enable the
transaction management system to choose a plan for
a goal (G in Fig. 4) and to determine the constraints
({c
1
,..,c
4
} in Fig. 5) for each sub-goal in that plan.
A plan’s performance estimate is the expected time
“t” and cost “c” to satisfy the plan’s goal. These
estimates will be calculated from performance
estimates for each atomic sub-system. The
parameters t and c are assumed to be normally
distributed—this is a wild assumption—but it
provides a framework for identifying measurements
that abnormal. Given a parameter, p, that is
assumed to be normally distributed, an estimate, µ
p
,
for the mean of p is revised on the basis of the i’th
observation ob
i
to µ
p
new
=
(1 - α) _ ob
i
+ α _ µ
p
old
which, given a starting
value µ
p
initial
, and some constant α, 0 < α < 1,
approximates the geometric mean \f(
\O(
Σ,
i=1
,
n
) α
n-i
_ ob
i
, \O(Σ,
i=1
,
n
) α
n-i
) of the set
of observations {ob
i
} where i = n is the most recent
observation. In the same way, an estimate, σ
p
, for
\r(\f(2,π)) times the standard deviation of p is
revised on the basis of the i’th observation ob
i
to
σ
p
new
= (1 – α) _ | ob
i
– µ
p
old
| + α _ σ
p
old
which, given a starting value σ
p
initial
, and some
constant α, 0 < α < 1, approximates the geometric
mean \f( \O(
Σ,
i=1
,
n
) α
n-i
_ | ob
i
–
µ
p
|,\O(Σ,
i=1
,
n
) α
n-i
) . The constant α is chosen
on the basis of the stability of the observations. For
example, if α = 0.85 then “everything more than
twenty trials ago” contributes less than 5% to the
weighted mean.
Given a transaction and its constraints
(expressed in terms of t and s), the transaction
management system makes two decisions. First it
selects a feasible plan for that transaction’s goal.
Second it determines the constraints on each sub-
goal in that plan. Then further plans are selected for
those sub-goals, and so on. Each time a plan for
goal G is used measurements are made of t and c for
each sub-goal in that plan. Further each of those
sub-goals may be invoked by other plans. So the
estimates of the mean and standard deviation of t
and c for those sub-goals may be expected to be
more accurate than the estimates for goal G. So
each time a plan is considered, the t and c estimates
for its goal are re-computed from those on the
estimated costliest path through the plan.
Plan A for goal [G, c] is feasible if
c > µ
A
+ κ _ σ
A
, where c is expressed in terms of t
and c, µ and σ are expressed likewise, and κ is a
constant usually > 1. If c < µ
A
– κ _ σ
A
then the
plan is not expected to achieve its goal within
constraint c. This enables the constraints to be
relaxed on each sub-goal so that the estimated
costliest path through the plan satisfies c. If a sub-
goal SG
i
of plan P for goal Gis not achieved within
its constraint c
i
then first another plan is sought for
SG
i
and for any other as-yet-unsatisfied ‘down
stream’ sub-goals, for which an allocation of
constraints in P is feasible, and second the whole
plan P fails and another plan is sought for G with
tighter constraints than c.
Given a goal G with constraints c the
transaction management system first identifies a set
of feasible plans for G. Then from this set the
system selects a plan for a given goal G using the
stochastic strategy: the probability that a plan is
selected is the probability that that plan is the “best”
plan. This strategy has been found to work well for
managing high level processes [6]. Here best may
mean “the most likely to satisfy the constraints on
G” or some other criterion such as “the plan likely to
deliver the best quality advice” as discussed below.
Given two plans A and B for the same goal G, if the
constraint on G is represented by a parameter p (in
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
328

terms of t and c) that is assumed to be normally
distributed then the probability that plan A is “better
than” plan B is the probability that (p
A
– p
B
) > 0.
Using elementary statistics, an estimate for this
probability is given by the area under the normal
distribution with:
mean = µ
A
– µ
B
[ where µ
A
and µ
B
are
estimates of the means of p
A
and p
B
]
standard
deviation = \r(\f(π,2) _ (σ\O(
2
,
A
) + σ\O(
2
,
B
))) [
where σ
A
and σ
B
are estimates of \r(\f(2,π)) times
the standard deviations of p
A
and p
B
] for x > 0.
This method may be extended to estimate the
probability that one plan is better than a number of
other plans.
The measurement of the quality, q, of work in
any business process is seldom available to the
management system except through subjective
assessment. This issue complicates the “optimal”
management of all business processes. In the
experimental e-market some information sources
may reasonably be given quality estimates. For
example, subjective estimates of the mean and
variance may be attached to text by a particular
journalist in a news feed. If these estimates are
available then the notion of “best” may be extended
to include quality. However, if “best” is to mean
some combination of q, c and t then these
parameters may need to be measured in the same
units, such as some monetary value.
The adjustment of estimates in the light of
measurements that fall outside expected ranges is
achieved using the geometric weighted mean
method used to estimate s and t. These multipliers
υ
ij
mean: if measurement m
i
of service i lies outside
the expected range then multiply the estimate for
service j by \F( m
i
,µ
i
) _ υ
ij
. This is crude but in
a sense these multipliers are no cruder than the
estimates that they are adjusting. What is known is
the network topology and so too potential causal
links between components’ performance. The use
of some form of belief net [8] is appealing in that the
learning mechanism has a scientific basis, although
here the nets will need to represent conditional t and
c estimates rather than conditional probabilities.
This is presently being investigated.
7 CONCLUSION
Two classes of business process are goal-driven
processes and knowledge-driven processes. Goal-
driven processes may be managed but they are
inherently unpredictable. The management of
knowledge-driven processes that involve human
agents is seldom feasible due to the size of the
process knowledge base. A significant class of
knowledge-driven processes is e-market transactions
in which the process knowledge base is the Internet.
These are managed using a multiagent system that is
supported by a suite of data and text mining bots
whose output is combined using a belief network.
The proactive component of these agents is specified
by plans. For goal-driven processes the proactive
‘succeed’ exit leads the way, and for knowledge-
driven processes the reactive ‘abort’ exit is used to
determine the process’ direction as knowledge is
revealed.
REFERENCES
Debenham, JK. Supporting the actors in an
electronic market place. In proceedings Twenty
First International Conference on Knowledge
Based Systems and Applied Artificial
Intelligence, ES’2001: Applications and
Innovations in Expert Systems IX, Cambridge
UK, December 2001, pp29-42.
Fischer, L. (Ed). Workflow Handbook 2001. Future
Strategies, 2000.
van der Aalst, W. & van Hee, K. Workflow
Management: Models, Methods, and Systems.
MIT Press (2001).
Jennings, N.R., Faratin, P., Norman, T.J., O’Brien,
P. and Odgers, B. (2000) Autonomous Agents
for Business Process Management. Int. Journal
of Applied Artificial Intelligence 14 (2) 145-189.
Rao, A.S. and Georgeff, M.P. “BDI Agents: From
Theory to Practice”, in proceedings First
International Conference on Multi-Agent
Systems (ICMAS-95), San Francisco, USA, pp
312—319.
Debenham, JK. Supporting knowledge-driven
processes in a multiagent process management
system. In proceedings Twentieth International
Conference on Knowledge Based Systems and
Applied Artificial Intelligence, ES’2000:
Research and Development in Intelligent
Systems XV, Cambridge UK, December 2000,
pp273-286.
MANAGING E-MARKET TRANSACTION PROCESSES: Exploring the limits of process management with a multiagent
system
329

Debenham, JK and Simoff, S. Investigating the
Evolution of Electronic Markets. In proceedings
Sixth International Conference on Cooperative
Information Systems, CoopIS 2001, Trento,
Italy, September 5-7, 2001, pp344-355.
Cowell, RG, Dawid, AP, Lauritzen, SL and
Spiegelhater, DJ. Probabilistic Networks and
Expert Systems. Springer-Verlag, (1999)
Weiss, G. (ed) (1999). Multi-Agent Systems. The
MIT Press: Cambridge, MA.
Klemperer, P. (Ed). The Economic Theory Of
Auctions. Edward Elgar Publishing (2000).
Debenham, JK and Simoff, S. Designing a Curious
Negotiator. In proceedings Third International
Workshop on Negotiations in electronic markets
- beyond price discovery - e-Negotiations 2002,
September 2002, Aix-en-Provence, France.
Watkins, M. Breakthrough Business Negotiation—A
Toolbox for Managers. Jossey-Bass, 2002.
Myerson, R. & Satterthwaite, M. Efficient
Mechanisms for Bilateral Trading. Journal of
Economic Theory, 29, 1–21, April 1983.
Kraus, S. Strategic Negotiation in Multiagent
Environments. MIT Press, 2001.
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
330