AN AGENT FOR EMERGENT PROCESS MANAGEMENT
John Debenham
University of Technology, Sydney
PO Box 123, Broadway, NSW 2007, Australia
Keywords:
Intelligent Agents, Business Process, Agents for Internet Computing.
Abstract:
Emergent processes are business processes whose execution is determined by the prior knowledge of the agents
involved and by the knowledge that emerges during a process instance. 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 trading
orders, requests for advice and information, are knowledge-driven processes for which the knowledge base is
the Internet, and so representing the knowledge is not at issue. Multiagent systems are an established platform
for managing complex business processes. What is needed for emergent process management is an intelligent
agent that is driven not by a process goal, but by an in-flow of knowledge, where each chunk of knowledge
may be uncertain. These agents should assess the extent to which it chooses to believe that the information is
correct, and so they require an inference mechanism that can cope with information of differing integrity. An
agent is described that achieves this by using ideas from information theory, and by using maximum entropy
logic to derive integrity estimates for knowledge about which it is uncertain. Emergent processes are managed
by these agents that extract the process knowledge from this knowledge base — the Internet — using a suite
of data mining bots. The agents make no assumptions about the internals of the other agents in the system
including their motivations, logic, and whether they are conscious of a utility function. These agents focus
only on the information in the signals that they receive.
1 INTRODUCTION
Emergent processes are business processes that are
not predefined and are ad hoc. These processes typi-
cally take place at the higher levels of organisations
(Dourish, 1998), and are distinct from production
workflows (Fischer, 2003). Emergent processes are
opportunistic in nature whereas production workflows
are routine. How an emergent process will terminate
may not be known until the process is well advanced.
The tasks involved in an emergent process are typi-
cally not predefined and emerge as the process devel-
ops. Those tasks may be carried out by collaborative
groups as well as by individuals (Smith and Fingar,
2003) and may involve informal meetings, business
lunches and so on. For example, in an e-market con-
text an emergent process could be triggered by “lets
try to establish a business presence in Hong Kong”.
Further, the goal of an emergent process instance may
mutate as the instance matures. So unlike “lower-
order” processes, the goal of an emergent process in-
stance may not be used as a focus for the management
of that instance.
Emergent processes contain “knowledge-driven”
sub-processes, but may also contain conventional
“goal-driven” sub-processes. A knowledge-driven
process is guided by its “process knowledge” and
“performance knowledge”. The goal of a knowledge-
driven process may not be fixed and may mutate.
On the other hand, the management of a goal-driven
process instance is guided by its goal which is fixed.
A multiagent system to manage the “goal-driven”
processes is described in (Debenham, 2000). In that
system each human user is assisted by an agent which
is based on a generic three-layer, BDI hybrid agent ar-
chitecture. The term individual refers to a user/agent
pair. The general business of managing knowledge-
driven processes is illustrated in Fig. 1, and will be
discussed in Sec. 2.
Process management is an established application
3
Debenham J. (2005).
AN AGENT FOR EMERGENT PROCESS MANAGEMENT.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 3-10
DOI: 10.5220/0002519400030010
Copyright
c
 SciTePress

area for multi-agent systems (Singh, 2004) although
emergent processes are typically handled either man-
ually or by CSCW systems rather than by process
management systems. The use of these two technolo-
gies is not elegant and presents a barrier to a unified
view of emergent process management.
In an experimental e-market, transactions 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 mar-
ket data as well as from news feeds and other Inter-
net data. In this e-market every market transaction
is managed as a business process. To achieve this,
suitable process management machinery has been de-
veloped. To investigate what is “suitable” the essen-
tial features of these transactions are related to two
classes of process that are at the “high end” of process
management feasibility (van der Aalst and van Hee,
2002). The two classes are goal-driven processes
and knowledge-driven processes — Sec. 2. The term
“business process management” is generally used to
refer to the simpler class of workflow processes (Fis-
cher, 2003), although there are notable exceptions us-
ing multiagent systems (Singh, 2004).
The agent architecture described extends the sim-
ple, offer-exchange, bargaining agent described in
(Debenham, 2004). The agent described here is
driven by the contents of a knowledge base that rep-
resents the agent’s world model in probabilistic first-
order logic, and manages emergent processes. Each
message that the agent receives from another agent
reveals valuable information about the sender agent’s
position. The agent aims to respond with messages
that have comparable information revelation. In this
way it aims to gain the trust of its opponent. The agent
does not necessarily strive to optimize its utility and
aims to make informed decisions in an information-
rich but uncertain environment.
The emergent process management agent, Π, at-
tempts to fuse the agent interaction with the informa-
tion that is generated both by and because of it. To
achieve this, it draws on ideas from information the-
ory rather than game theory. Π decides what to do —
such as what message to send — on the basis of its
information that may be qualified by expressions of
degrees of belief. Π uses this information to calculate,
and continually re-calculate, probability distributions
for that which it does not know. One such distribution,
over the set of all possible actions, expresses Π’s be-
lief in the suitability to herself of the system perform-
ing that action. Other distributions attempt to predict
the behavior of its opponent, Ω say, — such as what
proposals she might accept, and of other unknowns
that may effect the process outcome. Π makes no as-
sumptions about the internals of the other agents in
the system, including whether they have, or is even
aware of the concept of, utility functions. Π is purely
concerned with the other agents’ behaviors — what
they do — and not with assumptions about their mo-
tivations. This somewhat detached stance is appropri-
ate for emergent process management in which each
agent represents the interests of it owner, whilst at the
same time attempting to achieve the social goal of
driving the processes towards a satisfactory conclu-
sion.
As with the agent described in (Debenham, 2004),
the process management agent described here does
not assume that it has a von Neumann-Morgerstern
utility function. The agent makes assumptions about:
the way in which the integrity of information will de-
cay, and some of the preferences that its opponent
may have for some deals over others. It also assumes
that unknown probabilities can be inferred using max-
imum entropy inference (MacKay, 2003), ME, which
is based on random worlds (Halpern, 2003). The max-
imum entropy probability distribution is “the least bi-
ased estimate possible on the given information; i.e. it
is maximally noncommittal with regard to missing in-
formation” (Jaynes, 1957). In the absence of knowl-
edge about the other agents’ decision-making appa-
ratuses the process management agent assumes that
the “maximally noncommittal” model is the correct
model on which to base its reasoning.
2 PROCESS MANAGEMENT
Following (Fischer, 2003) a business process is “a
set of one or more linked procedures or activities
which collectively realise a business objective or pol-
icy goal, normally within the context of an organi-
sational structure defining functional roles and rela-
tionships”. Implicit in this definition is the idea that
a process may be repeatedly decomposed into linked
sub-processes until those sub-processes are activities
which are atomic pieces of work. [viz (Fischer, 2003)
“An activity is a description of a piece of work that
forms one logical step within a process.”].
A particular process is called a (process) instance.
An instance may require that certain things should be
done; such things are called tasks. A trigger is an
event that leads to the creation of an instance. The
goal of an instance is a state that the instance is try-
ing to achieve. The termination condition of an in-
stance is a condition which if satisfied during the life
of an instance causes that instance to be destroyed
whether its goal has been achieved or not. The patron
of an instance is the individual who is responsible for
managing the life of that instance. At any time in a
process instance’s life, the history of that instance is
the sequence of prior sub-goals and the prior sequence
of knowledge inputs to the instance. The history is
“knowledge of all that has happened already”.
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
4

Three classes of business process are defined in
terms of their management properties (ie: in terms of
how they may be managed).
• A task-driven process has a unique decomposition
into a — possibly conditional — sequence of ac-
tivities. Each of these activities has a goal and
is associated with a task that “always” achieves
this goal. Production workflows are typically task-
driven processes.
• A goal-driven process has a process goal, and
achievement of that goal is the termination condi-
tion for the process. The process goal may have
various decompositions into sequences of sub-
goals where these sub-goals are associated with
(atomic) activities and so with tasks. Some of these
sequences of tasks may work better than others, and
there may be no way of knowing which is which
(Smith and Fingar, 2003). A task for an activity
may fail outright, or may be otherwise ineffective
at achieving its goal. In other words, failure is a
feature of goal-driven processes. If a task fails then
another way to achieve the process goal may be
sought.
• A knowledge-driven process may have a process
goal, but the goal may be vague and may mutate
(Dourish, 1998). Mutations are determined by the
process patron, often in the light of knowledge gen-
erated during the process. At each stage in the per-
formance of a knowledge-driven process the “next
goal” is chosen by the process patron; this choice
is made using general knowledge about the con-
text of the process — called the process knowl-
edge. The process patron also chooses the tasks
to achieve that next goal; this choice may be made
using general knowledge about the effectiveness of
tasks — called the performance knowledge. So in
so far as the process goal gives direction to goal-
driven — and task-driven — processes, the process
knowledge gives direction to knowledge-driven
processes. The management of knowledge-driven
processes is considerably more complex than the
other two classes of process. But, knowledge-
driven processes are “not all bad” — they typically
have goal-driven sub-processes which may be han-
dled in conventional way. A simplified view of
knowledge-driven process management is shown in
Fig. 1.
Managing knowledge-driven processes is rather
more difficult than goal-driven processes, see Fig. 1.
The complete representation, never mind the mainte-
nance, of the process knowledge may be an enormous
job. But the capture of at least some of the knowledge
generated during a process instance may not be diffi-
cult if the tasks chosen used virtual documents such
as workspace technology, for example. Some perfor-
mance knowledge is not difficult to capture, represent
Process Goal
(what we presently
think we are trying
to achieve over all)
Performance Knowledge
(knowledge of how
effective tasks are at
achieving things)
Process
Knowledge
(knowledge of
what has been
achieved so far)
Now-Goal
(what to try to
achieve next)
Activity
(what to do next)
Decompose
(in the context of the
process knowledge)
Do it —
(until
termination
condition
satisfied )
New
Activity
Knowledge
New
Process
Knowledge
Add to
Revise
Select
Add to
Figure 1: Knowledge-driven process management
and maintain. For example, measurements of how
long another agent took to complete a sub-process can
be very useful. So in the system described here, the
process knowledge is left in the heads of the patron
or nominated delegates, and the performance knowl-
edge is captured by the system. The initial selection
of the process goal is performed by the patron, and
so this action is completely unsupported by the sys-
tem, see Fig. 1. The possible subsequent mutation
of the process goal is performed by the agent using
the process knowledge, see Fig. 1. Task selection is
supported by the agent for e-market processes which
can, for example, be given authority to withdraw a
bid from two separate auctions and to negotiate for
a package of goods from a single supplier. In this
way the system provides considerable assistance in
the management of knowledge-driven processes. Fur-
ther, if a now-goal is associated with a goal-driven, or
task-driven, sub-process then the management system
is given full responsibility for the management of that
sub-process.
3 EMERGENT PROCESS AGENT
Π operates in an information-rich environment that
includes the Internet. The integrity of Π’s informa-
tion, including information extracted from the Inter-
net, will decay in time. The way in which this decay
occurs will depend on the type of information, and
on the source from which it is drawn. Little appears
to be known about how the integrity of real informa-
AN AGENT FOR EMERGENT PROCESS MANAGEMENT
5

tion, such as news-feeds, decays, although the effect
of declining integrity has been analyzed. For exam-
ple, (Bernhardt and Miao, 2004) considers how de-
lays in the acquisition of trading data effect trading
outcomes.
One source of Π’s information is the signals re-
ceived from Ω. These include offers from Ω to Π,
the acceptance or rejection by Ω of Π’s offers, and
claims that Ω sends to Π. This information is aug-
mented with sentence probabilities that represent the
strength of Π’s belief in its truth. If Ω rejected Π’s
offer of $8 two days ago then what is Π’s belief now
in the proposition that Ω will accept another offer of
$8 now? Perhaps it is around 0.1. A linear model is
used to model the integrity decay of these beliefs, and
when the probability of a decaying belief approaches
0.5
1
the belief is discarded. The model of decay could
be exponential, quadratic or what ever.
3.1 Interaction Protocol
A deal is a pair of commitments δ
Π:Ω
(π, ω) between
an agent Π and an opponent agent Ω, where π is Π’s
commitment and ω is Ω’s commitment. D = {δ
i
}
D
i=1
is the deal set — ie: the set of all possible deals. If the
discussion is from Π’s point of view then the subscript
“Π : Ω” may be omitted. These commitments may
involve multiple issues and not simply a single issue
such as trading price. The set of terms, T , is the set of
all possible commitments that could occur in deals in
the deal set. An agent may have a real-valued utility
function: U : T → ℜ, that induces an ordering on
T . For such an agent, for any deal δ = (π, ω) the
expression U(ω) − U(π) is called the surplus of δ,
and is denoted by L(δ) where L : T × T → ℜ. For
example, the values of the function U may expressed
in units of money. It may not be possible to specify
the utility function either precisely or with certainty.
2
This is addressed in Sec. 4 where a predicate ΩAcc(.)
represents the acceptability of a deal to Ω.
The agents communicate using sentences in a first-
order language C. This includes the exchange, accep-
tance and rejection of offers. C usual trading pred-
icates including the following: Offer(δ), Accept(δ),
Reject(δ), Bid(δ) and Quit(.), where Offer(δ) means
“the sender is offering you a deal δ”, Accept(δ) means
“the sender accepts your deal δ”, Reject(δ) means
“the sender rejects your deal δ”, Bid(δ) means “the
sender submits the bid δ” and Quit(.) means “the
sender quits — the negotiation ends”.
1
A sentence probability of 0.5 represents null informa-
tion, ie: “maybe, maybe not”.
2
The often-quoted oxymoron “I paid too much for it,
but its worth it.” attributed to Samuel Goldwyn, movie pro-
ducer, illustrates that intelligent agents may negotiate with
uncertain utility.
3.2 Agent Architecture
Π uses the language C for external communication,
and the language L for internal representation. Two
predicates in L are: ΠAcc(.) and ΩAcc(.). The propo-
sition (ΠAcc(δ) | I
t
) means: “Π will be comfortable
accepting the deal δ given that Π knows information
I
t
at time t”. The idea is that Π will accept deal δ if
P(ΠAcc(δ) | I
t
) ≥ α for some threshold constant α.
The precise meaning that Π gives to ΠAcc(.) is de-
scribed in Sec. 4. The proposition ΩAcc(δ) means “Ω
is prepared to accept deal δ”. The probability distrib-
ution P(ΩAcc(.)) is estimated in Sec. 5.
Each incoming message M from source S received
at time t is time-stamped and source-stamped, M
[S,t]
,
and placed in an in box, X , as it arrives. Π has an
information repository I, a knowledge base K and a
belief set B. Each of these three sets contains state-
ments in a first-order language L. I contains state-
ments in L together with sentence probability func-
tions of time. I
t
is the state of I at time t and may
be inconsistent. At some particular time t, K
t
con-
tains statements that Π believes are true at time t,
such as ∀x(Accept(x) ↔ ¬Reject(x)). The belief set
B
t
= {β
i
} contains statements that are each qualified
with a given sentence probability, B(β
i
), that repre-
sents Π’s belief in the truth of the statement at time
t. The distinction between the knowledge base K and
the belief set B is simply that K contains unqualified
statements and B contains statements that are quali-
fied with sentence probabilities. K and B play differ-
ent roles in the method described in Sec. 3.3; K
t
∪ B
t
is required by that method to be consistent.
Π’s actions are determined by its “strategy”. A
strategy is a function S : K × B → A where A
is the set of actions. At certain distinct times the
function S is applied to K and B and the agent does
something. The set of actions, A, includes send-
ing Offer(.), Accept(.), Reject(.), Quit(.) messages
and claims to Ω. The way in which S works is de-
scribed in Secs. 5. Two “instants of time” before the
S function is activated, an “import function” and a
“revision function” are activated. The import func-
tion I : (X × I
t
−
) → I
t
clears the in-box, using
its “import rules”. An import rule takes a message
M, written in language C, and from it derives sen-
tences written in language L to which it attaches de-
cay functions, and adds these sentences together with
their decay functions to I
t
−
to form I
t
. These de-
cay functions are functions of the message type, the
time the message arrived and the source from which it
came — an illustration is given below. An import rule
has the form: P(S | M
[Ω,t]
) = f (M, Ω, t) ∈ [0, 1],
where S is a statement, M is a message and f is
the decay function. Then the belief revision function
R : I
t
−
→ (I
t
× K
t
× B
t
) deletes any statements
in I
t
−
whose sentence probability functions have a
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
6

value that is ≈ 0.5 at time t. From the remaining
statements R selects a consistent set of statements
and instantiates their sentence probability functions
to time t, and places the unqualified statements from
that set in K
t
and the qualified statements, together
with their sentence probabilities, in B
t
.
An example now illustrates the ideas in the previ-
ous paragraph. Suppose that the predicate ΩAcc(δ)
means that “deal δ is acceptable to Ω”. Suppose
that Π is attempting to trade a good “g” for cash.
Then a deal δ(π, ω) will be δ(g, x) where x is an
amount of money. If Π assumes that Ω would pre-
fer to pay less than more then I
t
will contain: ι
0
:
(∀gxy)((x ≥ y) → (ΩAcc(g, x)) → ΩAcc(g, y)).
Suppose Π uses a simple linear decay for its import
rules: f(M, Ω, t
i
) = trust(Ω) + (0.5 − trust(Ω)) ×
t−t
i
decay(Ω)
, where trust(Ω) is a value in [0.5, 1] and
decay(Ω) > 0.
3
trust(Ω) is the probability attached
to S at time t = t
i
, and decay(Ω) is the time pe-
riod taken for P(S) to reach 0.5 when S is dis-
carded. Suppose at time t = 7, Π receives the
message: Offer(g, $20)
[Ω,7]
, and has the import rule:
P(ΩAcc(g, x) | Offer(g, x)
[Ω,t
i
]
) = 0.8 − 0.025 ×
(t − t
i
), ie: trust is 0.8 and decay is 12. Then, in
the absence of any other information, at time t = 11,
K
t
11
contains ι
0
and B
t
11
contains ΩAcc(g, $20) with
a sentence probability of 0.7.
Π uses three things to make offers: an estimate of
the likelihood that Ω will accept any offer [Sec. 5],
an estimate of the likelihood that Π will, in hind-
sight, feel comfortable accepting any particular offer
[Sec. 4], and an estimate of when Ω may quit and
leave the negotiation — see (Debenham, 2004). Π
supports its negotiation with claims with the aim of
either improving the outcome — reaching a more ben-
eficial deal — or improving the process — reaching a
deal in a more satisfactory way.
3.3 Random worlds
Let G be the set of all positive ground literals that can
be constructed using the predicate, function and con-
stant symbols in L. A possible world is a valuation
function V : G → {⊤, ⊥}. V denotes the set of all
possible worlds, and V
K
denotes the set of possible
worlds that are consistent with a knowledge base K
(Halpern, 2003).
A random world for K is a probability distribu-
tion W
K
= {p
i
} over V
K
= {V
i
}, where W
K
ex-
presses an agent’s degree of belief that each of the
3
In this example, the value for the probability is given by
a linear decay function that is independent of the message
type, and trust and decay are functions of Ω only. There is
scope for using learning techniques to refine the trust and
decay functions in the light of experience.
possible worlds is the actual world. The derived sen-
tence probability of any σ ∈ L, with respect to a ran-
dom world W
K
is (∀σ ∈ L):
P
W
K
(σ) ,
X
n
{ p
n
: σ is ⊤ in V
n
} (1)
A random world W
K
is consistent with the agent’s
beliefs B if: (∀β ∈ B)(B(β) = P
W
K
(β)). That
is, for each belief its derived sentence probability as
calculated using Eqn. 1 is equal to its given sentence
probability.
The entropy of a discrete random variable X with
probability mass function {p
i
} is (MacKay, 2003):
H(X) = −
P
n
p
n
log p
n
where: p
n
≥ 0 and
P
n
p
n
= 1. Let W
{K,B}
be the “maximum entropy
probability distribution over V
K
that is consistent with
B”. Given an agent with K and B, its derived sentence
probability for any sentence, σ ∈ L, is:
(∀σ ∈ L)P(σ) , P
W
{K,B}
(σ) (2)
Using Eqn. 2, the derived sentence probability for any
belief, β
i
, is equal to its given sentence probability. So
the term sentence probability is used without ambigu-
ity.
If X is a discrete random variable taking a fi-
nite number of possible values {x
i
} with probabili-
ties {p
i
} then the entropy is the average uncertainty
removed by discovering the true value of X, and is
given by H(X) = −
P
n
p
n
log p
n
. The direct op-
timization of H(X) subject to a number, θ, of linear
constraints of the form
P
n
p
n
g
k
(x
n
) =
g
k
for given
constants
g
k
, where k = 1, . . . , θ, is a difficult prob-
lem. Fortunately this problem has the same unique
solution as the maximum likelihood problem for the
Gibbs distribution (Pietra et al., 1997). The solution
to both problems is given by:
p
n
=
exp

−
P
θ
k=1
λ
k
g
k
(x
n
)

P
m
exp

−
P
θ
k=1
λ
k
g
k
(x
m
)

(3)
n = 1, 2, · · · where the constants {λ
i
} may be calcu-
lated using Eqn. 3 together with the three sets of con-
straints: p
n
≥ 0,
P
n
p
n
= 1 and
P
n
p
n
g
k
(x
n
) =
g
k
. The distribution in Eqn. 3 is known as Gibbs dis-
tribution.
4 SUITABILITY OF AN ACTION
The proposition (ΠAcc(δ) | I
t
) was introduced in
Sec. 3.2. This section describes how the agent esti-
mates its beliefs of whether this proposition is true
for various δ.
AN AGENT FOR EMERGENT PROCESS MANAGEMENT
7

4.1 An Exemplar Application
An exemplar application follows. Π is placing bids
in an e-market attempting to purchase of a particu-
lar second-hand motor vehicle, with some period of
warranty, for cash. So the two issues in this nego-
tiation are: the period of the warranty, and the cash
consideration. A deal δ consists of this pair of is-
sues, and the deal set has no natural ordering. Sup-
pose that Π wishes to apply ME to estimate values
for: P(ΩAcc(δ)) for various δ. Suppose that the
warranty period is simply 0, · · · , 4 years, and that
the cash amount for this car will certainly be at least
$5,000 with no warranty, and is unlikely to be more
than $7,000 with four year’s warranty. In what fol-
lows all price units are in thousands of dollars. Sup-
pose then that the deal set in this application consists
of 55 individual deals in the form of pairs of warranty
periods and price intervals: { (w, [5.0, 5.2)), (w, [5.2,
5.4)), (w, [5.4, 5.6)), (w, [5.6, 5.8), (w, [5.8, 6.0)),
(w, [6.0, 6.2)), (w, [6.2, 6.4)), (w, [6.4, 6.6)), (w,
[6.6, 6.8)), (w, [6.8, 7.0)), (w, [7.0, ∞)) }, where w =
0, · · · , 4. Suppose that Π has previously received two
offers from Ω. The first is to offer 6.0 with no war-
ranty, and the second to offer 6.9 with one year’s war-
ranty. Suppose Π believes that Ω still stands by these
two offers with probability 0.8. Then this leads to
two beliefs: β
1
: ΩAcc(0, [6.0, 6.2)); B(β
1
) = 0.8,
β
2
: ΩAcc(1, [6.8, 7.0)); B(β
2
) = 0.8. Following
the discussion above, before “switching on” ME, Π
should consider whether it believes that P(ΩAcc(δ))
is uniform over δ. If it does then it includes both β
1
and β
2
in B, and calculates W
{K,B}
that yields esti-
mates for P(ΩAcc(δ)) for all δ. If it does not then it
should include further knowledge in K and B. For ex-
ample, Π may believe that Ω is more likely to bid for a
greater warranty period the higher her bid price. If so,
then this is a multi-issue constraint, that is represented
in B, and is qualified with a sentence probability.
4.2 Estimation of Beliefs
Here, agent, Π, is attempting to buy a second-hand
motor vehicle with a specific period of warranty as
described in Sec. 4.1. This section describes how Π
estimates: P(ΠAcc(δ) | I
t
). This involves the intro-
duction of four predicates into the language L: Me(.),
Suited(.), Good(.) and Fair(.).
General information is extracted from the World
Wide Web using special purpose bots that import and
continually confirm information. These bots commu-
nicate with Π by delivering messages to Π’s in-box
X using predicates in the communication language
C in addition to those described in Sec. 3.1. These
predicates include IsGood(Γ, Ω, r), and IsFair(Γ,
δ, s) meaning respectively that “according to agent
Γ, agent Ω is a good person to deal with certainty
P(Me( δ )) P(Suited( ω )) P(Good( Ω )) P(Fair( δ ))
P(
Π Acc(δ) | I
t
)
¡ ¨¨ ¨
Internet Market data
Agent
Ω
Agent Π
K
t
B
t
I
t
X
J
t
Figure 2: Estimating Π’s beliefs
r”, and “according to agent Γ, δ is a fair market
deal with certainty s”. The continual in-flow of in-
formation is managed as described in (Debenham,
2003). As described in Sec. 3.2, import functions
are applied to convert these messages into beliefs.
For example: P(Good(Ω) | IsGood(Γ, Ω, r)
[Θ,t
i
]
) =
f(IsGood, r, Γ, t), where Good(Ω) is a predicate in
the agents internal language L meaning “Ω will be a
good agent to do business with”. Likewise, IsFair(.)
messages in C are imported to I as Fair(.) statements
in L, where Fair(δ) means “δ is generally considered
to be a fair deal at least”.
With the motor vehicle application in mind,
P(ΠAcc(δ) | I
t
) is derived from conditional proba-
bilities attached to four other propositions: Suited(ω),
Good(Ω), Fair(δ), and Me(δ), where Suited(ω) means
“terms ω are perfectly suited to Π’s needs”, and
Me(δ) means “on strictly subjective grounds, the deal
δ is acceptable to Π”. These four probabilities are:
P(Suited(ω) | I
t
), P(Good(Ω) | I
t
), P(Fair(δ) |
I
t
∪ {Suited(ω), Good(Ω)}) and P(Me(δ) | I
t
∪
{Suited(ω), Good(Ω)}). The last two of these four
probabilities factor out both the suitability of ω and
the appropriateness of the opponent Ω. The third cap-
tures the concept of “a fair market deal” and the fourth
a strictly subjective “what ω is worth to Π”. The
“Me(.)” proposition is closely related to the concept
of a private valuation in game theory. This derivation
of P(ΠAcc(δ) | I
t
) from the four other probabilities
may not be suitable for assessing other types of deal.
For example, in eProcurement some assessment of the
value of an on-going relationship with an opponent
may be a significant issue. Also, for some low-value
trades, the inclusion of Good(.) may not be required.
The whole “estimation of beliefs” apparatus is il-
lustrated in Fig. 2.
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
8

5 INTERACTION
Π engages in bilateral bargaining with its opponent
Ω. Π and Ω each exchange offers alternately at suc-
cessive discrete times (Kraus, 2001). They enter into
a commitment if one of them accepts a standing offer.
The protocol has three stages:
1. Simultaneous, initial, binding offers from both
agents;
2. A sequence of alternating offers, and
3. An agent quits and walks away from the negotia-
tion.
In the first stage, the agents simultaneously send Of-
fer(.) messages to each other that stand for the entire
negotiation. These initial offers are taken as limits
on the range of values that are considered possible.
This is crucial to the method described in Sec. 3.3
where there are domains that would otherwise be un-
bounded. The exchange of initial offers “stakes out
the turf” on which the subsequent negotiation will
take place. In the second stage, an Offer(.) message is
interpreted as an implicit rejection, Reject(.), of the
opponent’s offer on the table. Second stage offers
stand only if accepted by return — Π interprets these
offers as indications of Ω’s willingness to accept —
they are represented as beliefs with sentence proba-
bilities that decay in time. The negotiation ceases ei-
ther in the second round if one of the agents accepts a
standing offer or in the final round if one agent quits
and the negotiation breaks down.
To support the offer-exchange process, Π has do
two different things. First, it must respond to offers
received from Ω — that is described in Sec. 4. Sec-
ond, it must send offers, and possibly information,
to Ω. This section describes machinery for estimat-
ing the probabilities P(ΩAcc(δ)) where the predicate
ΩAcc(δ) means “Ω will accept Π’s offer δ”. In the
following, Π is attempting to purchase of a particular
second-hand motor vehicle, with some period of war-
ranty, for cash from Ω as described in Sec. 4.1. So a
deal δ will be represented by the pair (w, p) where w
is the period of warranty in years and $p is the price.
Π assumes the following two preference relations
for Ω, and K contains:
κ
11
:
∀x, y, z((x < y) → (ΩAcc(y, z) → ΩAcc(x, z)))
κ
12
:
∀x, y, z((x < y) → (ΩAcc(z, x) → ΩAcc(z, y)))
As in Sec. 4, these sentences conveniently reduce the
number of possible worlds. The two preference rela-
tions κ
11
and κ
12
induce a partial ordering on the sen-
tence probabilities in the P(ΩAcc(w, p)) array from
the top-left where the probabilities are ≈ 1, to the
bottom-right where the probabilities are ≈ 0. There
are fifty-one possible worlds that are consistent with
K.
Suppose that the offer exchange has proceeded as
follows: Ω asked for $6,900 with one year war-
ranty and Π refused, then Π offered $5,000 with
two years warranty and Ω refused, and then Ω asked
for $6,500 with three years warranty and Π re-
fused. Then at the next time step B contains: β
11
:
ΩAcc(3, [6.8, 7.0)), β
12
: ΩAcc(2, [5.0, 5.2)) and
β
13
: ΩAcc(1, [6.4, 6.6)), and with a 10% decay in
integrity for each time step: P(β
11
) = 0.7, P(β
12
) =
0.2 and P(β
13
) = 0.9
Eqn. 3 is used to calculate the distribution W
{K,B}
which shows that there are just five different probabil-
ities in it. The probability matrix for the proposition
ΩAcc(w, p) is:
p  w 0 1 2 3 4
[7.0, ∞) 0.9967 0.9607 0.8428 0.7066 0.3533
[6.8, 7.0) 0.9803 0.9476 0.8330 0.7000 0.3500
[6.6, 6.8) 0.9533 0.9238 0.8125 0.6828 0.3414
[6.4, 6.6) 0.9262 0.9000 0.7920 0.6655 0.3328
[6.2, 6.4) 0.8249 0.8019 0.7074 0.5945 0.2972
[6.0, 6.2) 0.7235 0.7039 0.6228 0.5234 0.2617
[5.8, 6.0) 0.6222 0.6058 0.5383 0.4523 0.2262
[5.6, 5.8) 0.5208 0.5077 0.4537 0.3813 0.1906
[5.4, 5.6) 0.4195 0.4096 0.3691 0.3102 0.1551
[5.2, 5.4) 0.3181 0.3116 0.2846 0.2391 0.1196
[5.0, 5.2) 0.2168 0.2135 0.2000 0.1681 0.0840
In this array, the derived sentence probabilities for the
three sentences in B are shown in bold type; they are
exactly their given values.
Π’s negotiation strategy is a function S : K × B →
A where A is the set of actions that send Offer(.),
Accept(.), Reject(.) and Quit(.) messages to Ω. If Π
sends Offer(.), Accept(.) or Reject(.) messages to Ω
then she is giving Ω information about herself. In an
infinite-horizon bargaining game where there is no in-
centive to trade now rather than later, a self-interested
agent will “sit and wait”, and do nothing except, per-
haps, to ask for information. The well known bar-
gaining response to an approach by an interested party
“Well make me an offer” illustrates how a shrewd bar-
gainer may behave in this situation.
An agent may be motivated to act for various rea-
sons — three are mentioned. First, if there are
costs involved in the bargaining process due either
to changes in the value of the negotiation object with
time or to the intrinsic cost of conducting the nego-
tiation itself. Second, if there is a risk of breakdown
caused by the opponent walking away from the bar-
gaining table. Third, if the agent is concerned with
establishing a sense of trust (Ramchurn et al., 2003)
with the opponent —this could be the case in the es-
tablishment of a business relationship. Of these three
reasons the last two are addressed here. The risk of
breakdown may be reduced, and a sense of trust may
be established, if the agent appears to its opponent
to be “approaching the negotiation in an even-handed
manner”. One dimension of “appearing to be even-
handed” is to be equitable with the value of informa-
AN AGENT FOR EMERGENT PROCESS MANAGEMENT
9

tion given to the opponent. Various bargaining strate-
gies, both with and without breakdown, are described
in (Debenham, 2004), but they do not address this
issue. A bargaining strategy is described here that
is founded on a principle of “equitable information
gain”. That is, Π attempts to respond to Ω’s messages
so that Ω’s expected information gain similar to that
which Π has received.
Π models Ω by observing her actions, and by rep-
resenting beliefs about her future actions in the prob-
ability distribution P(ΩAcc). Π measures the value
of information that it receives from Ω by the change
in the entropy of this distribution as a result of rep-
resenting that information in P(ΩAcc). More gener-
ally, Π measures the value of information received in
a message, µ, by the change in the entropy in its en-
tire representation, J
t
= K
t
∪ B
t
, as a result of the
receipt of that message; this is denoted by: ∆
µ
|J
Π
t
|,
where |J
Π
t
| denotes the value (as negative entropy)
of Π’s information in J at time t. Although both Π
and Ω will build their models of each other using the
same data, the observed information gain will depend
on the way in which each agent has represented this
information. To support its attempts to achieve “eq-
uitable information gain” Π assumes that Ω’s reason-
ing apparatus mirrors its own, and so is able to esti-
mate the change in Ω’s entropy as a result of send-
ing a message µ to Ω: ∆
µ
|J
Ω
t
|. Suppose that Π
receives a message µ = Offer(.) from Ω and ob-
serves an information gain of ∆
µ
|J
Π
t
|. Suppose that
Π wishes to reject this offer by sending a counter-
offer, Offer(δ), that will give Ω expected “equitable
information gain”. δ = {arg max
δ
P(ΠAcc(δ) |
I
t
) ≥ α | (∆
Offer(δ)
|J
Ω
t
| ≈ ∆
µ
|J
Π
t
|)}. That is
Π chooses the most acceptable deal to herself that
gives her opponent expected “equitable information
gain” provided that there is such a deal. If there is not
then Π chooses the best available compromise δ =
{arg max
δ
(∆
Offer(δ)
|J
Ω
t
|) | P(ΠAcc(δ) | I
t
) ≥ α}
provided there is such a deal. If there is not then Π
does nothing.
6 CONCLUSION
Emergent processes are business processes whose ex-
ecution is determined by the prior knowledge of the
agents involved and by the knowledge that emerges
during a process instance. The establishment of a
sense of trust (Ramchurn et al., 2003) contributes to
the establishment of business relationships and to pre-
venting breakdown in one-off negotiation. The agent
architecture is based on a first-order logic represen-
tation, and so is independent of the number of nego-
tiation issues, although only two-issue bargaining is
illustrated here. Emergent processes are managed by
these agents that extract the process knowledge from
the Internet using a suite of data mining bots.
REFERENCES
Bernhardt, D. and Miao, J. (2004). Informed trading when
information becomes stale. The Journal of Finance,
LIX(1).
Debenham, J. (2000). Supporting strategic process. In
Proceedings Fifth International Conference on The
Practical Application of Intelligent Agents and Multi-
Agents, pages 237 – 256. Practical Applications Co.
Debenham, J. (2003). An eNegotiation Framework. In The
Twenty-Third International Conference on Innovative
Techniques and Applications of Artificial Intelligence,
AI’2003, pages 79 – 92. Springer Verlag.
Debenham, J. (2004). Bargaining with information. In
Jennings, N., Sierra, C., Sonenberg, L., and Tambe,
M., editors, Proceedings Third International Confer-
ence on Autonomous Agents and Multi Agent Systems
AAMAS-2004, pages 664 – 671. ACM.
Dourish, P. (1998). Using metalevel techniques in a flexi-
ble toolkit for CSCW applications. ACM Transactions
on Computer-Human Interaction (TOCHI), 5(2):109 –
155.
Fischer, L. (2003). The Workflow Handbook 2003. Future
Strategies Inc.
Halpern, J. (2003). Reasoning about Uncertainty. MIT
Press.
Jaynes, E. (1957). Information theory and statistical me-
chanics: Part I. Physical Review, 106:620 – 630.
Kraus, S. (2001). Strategic Negotiation in Multiagent Envi-
ronments. MIT Press.
MacKay, D. (2003). Information Theory, Inference and
Learning Algorithms. Cambridge University Press.
Pietra, S. D., Pietra, V. D., and Lafferty, J. (1997). Inducing
features of random fields. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 19(2):380–
393.
Ramchurn, S., Jennings, N., Sierra, C., and Godo, L.
(2003). A computational trust model for multi-agent
interactions based on confidence and reputation. In
Proceedings 5th Int. Workshop on Deception, Fraud
and Trust in Agent Societies.
Singh, M. (2004). Business Process Management: A Killer
Ap for Agents? In Jennings, N., Sierra, C., Sonen-
berg, L., and Tambe, M., editors, Proceedings Third
International Conference on Autonomous Agents and
Multi Agent Systems AAMAS-2004, pages 26 – 27.
ACM.
Smith, H. and Fingar, P. (2003). Business Process Manage-
ment (BPM): The Third Wave. Meghan-Kiffer Press.
van der Aalst, W. and van Hee, K. (2002). Workflow Man-
agement: Models, Methods, and Systems. The MIT
Press.
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
10