INTEGRATING AGENTS WITH CONNECTIONIST SYSTEMS
TO EXTRACT NEGOTIATION ROUTINES
Marisa Masvoula, Panagiotis Kanellis and Drakoulis Martakos
Department of Informatics and Telecommunications, Nationaland Kapodistrian University of Athens
University Campus, Athens15771, Greece
Keywords: Evolving Connectionist Systems, Negotiation Routines.
Abstract: Routinization is a technique of knowledge exploitation based on the repetition of acts. When applied to
negotiations it results the substitution of parts or even whole processes, disembarrassing negotiators from
significant deliberation and decision making effort. Although it has an important impact on negotiators, the
risk of establishing ineffective routines is evident. In our paper we discuss weaknesses and limitations and
we propose a generic framework to address them. We consider routines as evolving processes and we take
two orientations. The first concerns a communicative dimension to allow for external evaluation of the
applied routines and the second concerns enforcement of the system core with evolving structure that
adjusts to routine changes and flexibly incorporates new knowledge.
1 INTRODUCTION
Organizations mostly engage in improvement
strategies that concern knowledge exploitation and
exploration. Exploitation is considered similar to
recalling past experience to deal with current issues,
while exploration is considered similar to adopting
imitative or innovative tactics in the search for new
knowledge. Several techniques have been developed
to support exploitation and involve knowledge
refinement, routinization and elaboration of existing
ideas, paradigms, technologies, processes, strategies
and knowledge. Exploration, on the other hand, is
mostly supported through experimentation with new
ideas, paradigms, technologies, processes, strategies
and knowledge that improve on old ones. As
depicted in (Dyba, 2000) the basic balance problem
is “to undertake enough exploitation to ensure short
term results and concurrency to engage in
exploration to ensure long term survival”. Our study
mainly focuses on routinization as a means of
knowledge exploitation and its employment in
negotiations. Negotiation routines are understood as
repetitive acts in several stages of the negotiation
process. The identification of the context that
enables routine development and routine
‘subconscious’ retrieval, result to the development
of more effective negotiators who accomplish the
same outcomes with the minimum deliberation
effort. Nevertheless several limitations have been
acknowledged. Efficiency decrease that arises from
inflexibility and rigidity produced by routine
establishment, as well as lack of formal theoretical
frameworks and theories to routine application in
negotiations are among the most important
impediments. Furthermore routines should be
viewed as evolving processes, since the repetitive
acts may change over time even under similar
negotiation contexts, and distinct AI techniques are
not adequate to support such processes. In sections 2
and 3 we give a brief definition and background of
routines and their development in several
negotiation stages, as well as an analysis of
theoretical weaknesses and shortcomings of routine
establishment. In section 4 we identify the features
that a system acquiring and extracting negotiation
routines should be enhanced with, and we propose a
generic framework for such a purpose. In section 5
we explain how the proposal is expected to address
the discussed limitations and appose future research
issues.
251
Masvoula M., Kanellis P. and Martakos D. (2009).
INTEGRATING AGENTS WITH CONNECTIONIST SYSTEMS TO EXTRACT NEGOTIATION ROUTINES.
In Proceedings of the 11th International Conference on Enterprise Information Systems - Artificial Intelligence and Decision Support Systems, pages
251-256
DOI: 10.5220/0002155302510256
Copyright
c
SciTePress
2 ROUTINES FOR KNOWLEDGE
EXPLOITATION
“Routines” is the general term used for the definition
of all regular and predictable behavioral patterns
(Nelson, Winter, 1982). Although there is no final
agreement on an exact definition, researchers indent
that repetition of acts is a necessary precondition for
the acquisition and application of routines (Kesting,
2007). When an agent executes an act for the first
time, it faces the challenge of identifying and
evaluating alternatives that will lead to an intended
state. It is given the opportunity to observe the
results of the initial solution and if it is effective, it
can be applied to similar problems. The substance of
routines is the specific knowledge acquired by the
repeated planning and execution of an act combined
with the ability to apply this knowledge to specific
situations. It has the potential to substitute deliberate
planning and decision making since it is used to
determine what operations to conduct in order to
realize certain intended state. Routinization forces
agents to develop ‘best practices’.
2.1 Development of Routines in
Negotiation
Negotiation is most commonly defined as the search
for an agreement that satisfies the requirements of
two or more parties. Several scientific fields have
made contributions to the development of
negotiation theory. (Gulliver 1979, Robinson &
Volkov 1998, Putnam & Roloff 1992) discuss
negotiation models that follow normative,
prescriptive or descriptive approaches derived from
the application of economic theories, management
and social sciences respectively. Although
theoretical perspectives and definitions may vary,
there is a general consensus regarding the number of
stages in the negotiation process. Most approaches
model negotiation in three basic stages: (a)
prenegotiation stage basically concerned with
strategy formulation, commitment to rules,
observation of the other party behavior, definition of
issues and problem formulation (b) negotiation
dance, which is mainly concerned with the exchange
of offers and counter-offers and (c) execution of
negotiation results. Negotiations may be fully or
semi-automated processes and systems that facilitate
the decision making in several stages have been
developed. Nevertheless, the application of routines
in negotiation has not been extensively studied
(Kesting, Smolinski, 2007). In what context are
similar negotiation cases identified? How can
negotiation data (about products, shipping info,
particulars of buyers and sellers, orders and
dialogues of negotiation) be used to extract
knowledge which will be reused in further
negotiations? (Kesting, Smolinski, 2007) Introduces
a theoretical framework which identifies different
negotiation situations where routine can develop,
based on similarity and stability of negotiation
elements measured in two dimensions. The problem-
solving (substance of negotiation) dimension and the
communication dimension, concerning negotiation
partners. The framework in (Kesting, Smolinski,
2007) concludes that negotiation success is based on
a combination of problem-solving/analytic and
negotiation/communication skills that negotiators
possess, with routine application. Routine has an
important impact on negotiations, since it allows
negotiators to develop their capabilities, improve
their efficiency and save scarce planning capacities
and time by employing automated procedures.
3 PROBLEM STATEMENT
A commonly stated risk posed by routinization is the
application of ineffective acts. As stated in (Nelson
& Winter 1982, Winter 1986, Cohendet & Llerena
2003) with increasing repetitions, decision making
prior to the operation tends to decrease. The use of
routines entails rigidity and once a solution is
established, it is not further questioned. As a result,
formal routines alone are inadequate and might
demand improvisational skills. (Kesting, 2007)
Investigates the mechanisms that tend to make
routines resistant to change and are the major source
for inflexibility. The first concerns stability of
intention. Routines are developed to bring about a
specific intended state and as a consequence
intention defines a narrow frame that restricts the
extent of possible changes. The second mechanism
is derived from repetition which is prerequisite for
the application of routines. Repetition transcends
planning efforts, therefore routines are stable and do
not change. The third mechanism that is recognized
concerns automation. With increasing repetitions the
course of an act is automated, without further
deliberation.
The application of routines in negotiation has not
been researched, and the lack of formal models
describing the context and conditions that favor their
development is an obvious impediment. How the
negotiation environment is formally represented and
how is similarity of distinct contexts measured? To
ICEIS 2009 - International Conference on Enterprise Information Systems
252
overcome this theoretical limitation we make the
generic assumption that negotiation context can be
represented as a vector in n
1
dimensional space
(input space X), and similarity of distinct contexts is
calculated through distance vectors. Furthermore, we
assume that routines can be represented as vectors in
n
2
dimensional space (output space Y) and any
framework that results to routine extraction must
provide mappings from the input space X to the
output space Y.
Another important issue we need to address
concerns the decision of the most appropriate
technique that will provide mappings from input to
output space, as long as the development of a model
with a dynamic structure in order to accommodate
new knowledge and produce new routines in a
flexible and adaptive manner. Routinization must be
viewed as an evolving process, since the negotiation
context may change over time resulting to different
outcomes. A necessary first step is the study of
existing systems that exploit previous knowledge in
order to provide advising services during the various
phases of the negotiation process. A categorization
of such systems is provided in (Braun, Brzostowski
et al., 2005) and several AI methods and techniques
have been successfully developed and used.
Nevertheless a number of problems, which are
discussed in (Kasabov, 2007), arise from the
application of such techniques to evolving
processes:
1. “Difficulty in preselecting the system’s
architecture: Computational Intelligence models
usually retain a fixed architecture (e.g. number of
neurons and connections)”. This restricts us to a
view of a close problem space (fixed input and
output space), which is not the case if we
consider that negotiation environment and
routines may change over time.
2. “Catastrophic forgetting: Systems usually forget
a significant amount of old knowledge while
learning from new data.” This fact prohibits the
development of a framework that balances
knowledge exploitation with exploration.
3. “Excessive training time required: Training
usually requires many iterations of data
propagation through an ANN structure.” If
knowledge insertion and routine extraction is
time consuming the model is not considered
efficient.
4. “Lack of knowledge representation facilities:
Existing Computational Intelligence
architectures capture statistical parameters during
training, but do not facilitate the extraction of
evolving rules in terms of linguistically
meaningful information.” If knowledge is
represented in terms of IF-THEN rules (eg. If
<context> then <routine>) in order to be
linguistically meaningful, we can consider the
exchange of locutions between agents in terms of
exchange of experience.
4 PROPOSAL
The main problem posed by routinization is the risk
of acquiring and applying ineffective routines. To
this extent, we trust that the system producing
routines should be enhanced with learning
capabilities in order to evolve according to possible
satisfaction or dissatisfaction measures caused by
the applied routines. Furthermore it should be
enhanced with communicative abilities in order to
interact with the agents that apply the routines.
The underlying framework should combine
knowledge exploitation and knowledge exploration
techniques, and should have an evolving structure in
order to adapt to environmental changes. We outline
several desirable characteristics of an evolving
model that exploits and explores negotiation
knowledge.
1. Previous knowledge can be modeled in pairs of
data (x,y), where the desired output vector y is
known for an input vector x. In order to associate
data from the input space X to the output space
Y, the system must be enriched with supervised
learning abilities.
2. The system must interact with negotiators, in
order to trace dissatisfaction caused by the
application of a routine, request additional
planning and decision making and allow the
insertion of new knowledge to the system.
External evaluation by the negotiators results in
breaking the rigidity of routines. This feature
demands the development of a communication
protocol between the system and the negotiators,
as well as an adaptive structure to allow for
efficient knowledge insertion.
3. The system must be able to adapt to new data of
unknown distribution. We must take into account
that it may develop in its own machine learning
space M, different from the original data space Z.
New data vectors may have more or fewer
dimensions, resulting to dimensionality change
of the problem space over time. Therefore we
consider an open problem space.
INTEGRATING AGENTS WITH CONNECTIONIST SYSTEMS TO EXTRACT NEGOTIATION ROUTINES
253
4. System adaptation must be fast and not require
many iterations.
We propose a model that consolidates the above
characteristics and stems from the integration of an
intelligent agent and an Evolving Connectionist
System (ECOS). “An ECOS is an adaptive,
incremental learning and knowledge representation
system that evolves its structure and functionality,
where in the core of the system is a connectionist
architecture that consists of neurons and connections
between them” (Kasabov, 2007). We outline the
learning abilities of ECOS that position them as top
candidates for the combination of knowledge
exploitation and exploration.
1. “ECOS may evolve in open space, where the
dimensions of the space can change”
2. “They learn via incremental learning, possibly in
an online mode”
3. “They may learn continuously in a lifelong
learning mode
4. “They learn both as individual systems and as
evolutionary populations of such systems”
5. “They use constructive learning and have
evolving structures”
6. “They learn and partition the problem space
locally, thus allowing for a fast adaptation and
tracing the evolving processes over time”
7. “They evolve different types of knowledge
representation from data, mostly a combination
of memory-based, statistical and symbolic
knowledge”.
A simple ECOS system is eMLP (Kasabov, 2007),
which is a three-layer network with two layers of
connections. The first consists of input variables,
which represent the negotiation context. The second
contains rule nodes that evolve and allow for layer
growth during training. The rule nodes represent the
mapping of input to output space. These mappings
may be considered as IF-THEN rules, IF
<negotiation context> THEN <routine>.
The similarity of input vectors, which is a
precondition to routine application, is calculated
through the normalized distance between the input
vector and the incoming weight vector of the rule
node. Activation for the rule node is given through
A = 1 – Distance, which indicates that the more
similar the input vector (current negotiation context)
is to a previous case, activation tends to 1. A
sensitivity threshold may be used for the definition
of similarity. Supervised learning is also applied,
and the normalized output error is calculated. The
third layer represents the values of the output
variables, constituting the routines. The model
structure facilitates the accommodation of new
training examples within the evolving layer either by
modifying connection weights or by adding new
nodes. eMLP is suitable for online output space
expansion, because it tunes only the connection
weights of the local node and does not require
retraining of the whole system as in traditional
neural networks. In order to add a new node to the
output layer, the structure of the eMLP need to be
modified to accommodate the output node. The
modification affects only the output layer and its
connections with the evolving layer. As a result of
the training process new nodes will be added to the
evolving layer to represent new input-output
associations. The architecture of eMLP is illustrated
in figure 1, indicating the growing structure of the
evolving and output layer.
Figure 1: The evolving structure of eMLP (reproduced
with permission).
In order to meet all of the desirable
characteristics described above, the system must be
enriched with the ability to communicate with the
negotiators and receive external utility measurement
or even be provided with new information, evolving
the established routines. Our suggestion concerns the
integration of the eMLP with an intelligent agent,
which will act as a mediator providing negotiation
routines to the negotiator. The mediator will collect
previous negotiation knowledge from the negotiator
and train the eMLP which will be the agent core.
The communication protocol will support requests
for routine acquisition generated by the negotiator,
and routine proposals by the mediator. After the
application of each routine, the negotiator will
proceed to evaluation (by using a utility function)
and in case of dissatisfaction, the routine will break
and the negotiator will be challenged to apply his
skills for knowledge exploration. The new
knowledge will be provided back to the mediator
and result to the evolution of the eMLP.
ICEIS 2009 - International Conference on Enterprise Information Systems
254
5 EXPECTED RESULTS
Our research attempts to shed light to the field of
negotiation optimization, by focusing on the
acquisition and application of negotiation routines.
It is commonly stated that the main problem of
routinization is the risk of applying ineffective
routines. As discussed earlier, among the causes that
produce inflexible routines lays a repetitive
mechanism that prohibits further questioning or
evaluation of the acts. We trust that our framework
makes contribution to theory of negotiation routines
by presenting two essential directions for breaking
routine rigidity. The first concerns the identification
of the need of external evaluation after each routine
has been applied and the second concerns the
enrichment of the system core with an evolving
structure that adjusts its functionality to
accommodate new knowledge. The first direction
actually indicates the necessity to keep an open
communication channel with the environment, and
allow for the notification and triggering of routine
change. The second direction concerns the
identification of an appropriate structure that
embraces evolving processes, in order to absorb new
knowledge and proceed to routine change.
To this extent we have outlined four desirable
characteristics in section 4 that are met in our
proposed framework. The first concerns a hypothesis
of the representation of negotiation context and
routines as vector pairs. The second identifies the
need of external routine evaluation and dynamic
system structure. Our framework suggests the
integration of the dynamic system structure (ECOS)
with a negotiation agent, in order to allow the
interaction of the model that captures knowledge and
extracts routines with the negotiators, and be
notified in terms of a utility measure (after routine
application). The mediator agent will be motivated,
if the utility decreases, to request further planning by
the negotiators. The results (new knowledge) will
then be inserted to the dynamic structure. The third
and fourth characteristics concern the system’s
ability to develop in open space and quickly adapt to
the environment. These characteristics are met by
the ECOS structure, since it learns and partitions the
problem space locally. Furthermore in section 3 we
have identified several limitations posed by the
application of AI techniques in evolving processes.
These are addressed by ECOS algorithms, since they
apply fast one-pass learning (adaptation) and are
resistant to catastrophic forgetting (Kasabov, 2007).
Their structure is simple and grows in terms of
adaptation to the environment (the eMLP grows by
the addition of new nodes to the evolving and output
layer). Finally ECOS systems evolve different types
of knowledge representation from data; therefore our
representational hypothesis is not limited.
This framework is completely new and will be
tested in several stages of negotiation to provide
support to negotiators by the extraction of routines.
We trust that since the system is evolving it will lead
to efficient results in prenegotiation processes
(strategy formulation, commitment to rules,
opponent observation, issues and problem
formulation, other prenegotiation convensions), as
well as in negotiation processes (decision of
proposal and negotiation locutions). We have made
the hypothesis that knowledge can be provided to
the system in pairs of x, y vectors therefore the
model is generic and can be used for the substitution
of parts or even whole negotiations. Our future
research concerns extensive tracing of negotiation
repetitive acts, and the development and application
of our model in those that can be modeled as vector
pairs. If in several cases uncertainty is introduced to
vector dimensions and we need to address the issue
of missing values, the framework can be extended to
contain evolving fuzzy neural network (EFuNN)
instead of simple eMPL in its dynamic structure.
EFuNNs, function as eMLPs but have two extra
connection layers that represent fuzzy input and
output spaces. For these cases we will investigate the
integration of EFuNNs to our mediator framework.
Model validation, in terms of the generalization
ability of the ECOS core to produce good results on
new, unseen data samples, can be implemented by
splitting the original dataset to train and test sets and
calculating the actual error of the system. The most
commonly stated validation methods are simple train
and test split of data, k-fold cross validation and
leave-one-out cross validation. Furthermore, the
proposed structure itself suggests external evaluation
of the applied routines by the negotiators, which
contributes in measuring the overall system
performance.
REFERENCES
Braun, P., Brzostowski, J., Kersten, G., Baek Kim, J.,
Kowalczyk, R., Strecker, R & Vahidov, R 2005, ‘E-
Negotiation Systems and Software Agents: Methods,
Models and Applications’, in J. Gupta, G. Forgionne
& M. Mora (eds), Intelligent Decision-Making
Support Systems: Foundation, Applications, and
Challenges, Springer, London.
INTEGRATING AGENTS WITH CONNECTIONIST SYSTEMS TO EXTRACT NEGOTIATION ROUTINES
255
Cohendet, P & Llerena, P 2003, ‘Routines and Incentives:
The role of Communities in the Firm’, Industrial and
Corporate Change, vol. 12, no. 2, pp. 271-297
Dyba, T 2000, ‘Improvisation in Small Software
Organizations’, IEEE Software, vol. 17, no. 5, pp. 82-
87.
Gulliver, P 1979, Disputes and Negotiations: A Cross-
Cultural Perspective, Academic Press, New York.
Kasabov, N 2007, Evolving Connectionist Systems: The
Knowledge Engineering Approach, Springer-Verlag,
London, U.K.
Kesting, P 2007, ‘Stability and Changeability of Routines:
Mechanisms, Opportunities and Obstacles’, viewed 10
February 2009, <http://ssrn.com/abstract=887081>
Kesting, P & Smolinski, R 2007, ‘When Negotiations
become Routine: Not Reinventing the Wheel while
Thinking outside the Box’, Negotiation Journal, vol.
23, no. 4, pp. 419-438
Nelson, R.R & Winter, S. G 1982, An Evolutionary theory
of economic change, The Belknap Press of Harvard
University Press, Cambridge, MA and London.
Putnam, L., & Roloff, M 1992, ‘Communication
Perspectives on Negotiation’, in L. Putnam & M.
Roloff (eds), Communication and Negotiation (pp. 1-
17), Sage, Newbury Park, CA.
Robinson, W. N & Volkov, V 1998, ‘Supporting the
negotiation life cycle’, Communications of the ACM,
vol. 41, no. 5, pp. 95-102.
Winter, S.G 1986, ‘The Research Program of the
Behavioral Theory of the Firm: Orthodox Critique and
Evolutionary Perspective’, in B. Gilad & S. Kaish,
(eds), Handbook of Behavioral Economic, (vol A, pp.
151-188), JAI Press, Greenwich
ICEIS 2009 - International Conference on Enterprise Information Systems
256