COLLABORATION AMONG COMPETING FIRMS
An Application Model about Decision Making Strategy
Marco Remondino and Marco Pironti
e-business L@B, University of Turin, Torino, Italy
Keywords: Collaboration, Bias, Ego biased learning.
Abstract: To understand the adoption of collaborative systems, it is of great importance to know about economical
effects of collaboration itself. Decision makers should be able to evaluate potential drawbacks and
advantages of collaboration: strategies may be seen as a mixture of cost reduction, product differentiation
and improvement of decision making and/or planning. In this context information technology may help a
firm to create sustaining competitive advantages over competitors. It is less clear whether collaboration is of
any use in such an environment. According to the Economics literature, the most important factors affecting
benefits of collaboration are market structure, kind and degree of uncertainty faced by the firms, their risk
preferences and the collaboration propensity. The results depend on the way these factors are combined. We
present a microeconomic model and use techniques from game theory for the analysis. The way the model
is constructed will allow the derivation of closed-form solutions. Traditional learning models can't represent
individualities in a social system, or else they represented all of them in the same way – i.e.: as focused and
rational agents; they don’t represent individual inclinations and preferences. Results indicating whether
collaboration in various areas makes sense will be obtained. This makes it possible to judge the potential of
available collaborative technology. The basic presented model may be extended in various ways.
1 INTRODUCTION
Decision makers need to understand the economics
of collaboration in order to be able to evaluate the
potential of collaborative technology. Collaboration
among different actors may occur within a firm’s
boundary or across it. The economic effects of
collaboration between firms located along different
phases of the value chain (typically supplier-
purchaser-relationships) have been extensively
studied in the literature.
Usually, transaction cost theory is applied to
derive the “optimal” institutional structure
(Williamson, 1975). Basic institutional arrangements
are hierarchy, market and network cooperation
(Clemons et al., 1993). The use of collaborative
technology may be especially useful in case of
network cooperation. As Clemons et al. point out,
the use of IT triggers a move towards such
cooperation. As is well-known, strategies of firms
may be seen as a mixture of cost reduction, product
differentiation and improvement of decision making
and/or planning. Information technology may help a
firm to create sustaining competitive advantages
over competitors. Social models are not all and only
about coordination, like iterated games, and agents
could have a bias towards a particular behavior,
preferring it even if that’s not the best of the possible
ones. An example from the real world could be the
adoption of a technological innovation in a
company: even though it can be good for the
enterprise to adopt it, the managerial board could be
biased and could have a bad attitude towards
technology, perceiving a risk which is higher than
the real one. Thus, even by looking at the positive
figures coming from market studies and so on, they
could decide not to adopt it. This is something which
is not taken into consideration by traditional learning
methods, but that should be considered when
modeling social systems, where agents are often
supposed to mimic some human behavior. In order
to introduce these factors, a formal method is
presented in the paper: Ego Biased Learning (EBL).
A first result of the analysis shows that
maximization of expected utility may lead to
different optimal actions than maximization of
expected profits. While the latter in general is a
simple optimization problem, maximization of
expected utility requires knowledge of the utility
387
Remondino M. and Pironti M.
COLLABORATION AMONG COMPETING FIRMS - An Application Model about Decision Making Strategy.
DOI: 10.5220/0002796903870391
In Proceedings of the 6th International Conference on Web Information Systems and Technology (WEBIST 2010), page
ISBN: 978-989-674-025-2
Copyright
c
2010 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
function of the decision maker. Note, that this is a
more difficult and complex problem.
If costs of information sharing are sufficiently
low, information sharing generally is beneficial if
development know how is equally distributed. This
is an expected result since then development costs
may be reduced. This result, however, changes
significantly if know how for development is not
equally distributed between the competitors. In this
case, situations that are well-known from the
treatment of “prisoner-dilemma-situations” occur.
Results then strongly depend on the degree of risk
aversion and the market structure and nearly all
“prisoner-dilemma-situations” may be constructed
by suitably choosing the model parameters. First
mover and follower strategies are then optimal
choices depending on risk aversion of the decision
makers and market structure.
In some instances, the results obtained will be
surprising and contradict rational expectations.
It can be shown, e.g., that the use of information
to reduce uncertainty may be harmful for an
enterprise. Firms are paid for taking risks. If they try
to reduce such risks profits and expected utility may
decrease (even if risk is reduced at zero costs, see
e.g. Palfrey; 1982).
Once again, such surprising results show the
importance of understanding the economic effects of
collaboration before deciding on investments in
collaboration technologies. But it is very important
to consider collaboration expectation as a
predominant behavior of other firms.
2 COLLABORATIVE CULTURE
Generally, collaboration among competing firms
may occur in many ways. Some examples are joint
use of complex technological or marketing
processes, bundling products or setting standards.
Collaboration typically requires sharing information
and know how, as well as resources.
In the literature collaboration problems are
usually studied with the help of methods from
microeconomics and game theory. It turns out that
the most important factors affecting the usefulness
of collaboration are the following ones:
- Market Structure. If perfect competition
prevails collaboration is of limited use. No single
firm or proper subsets of firms may influence
market prices and/or quantities. In a monopolistic
environment there obviously is no room for
collaboration. Consequently, the interesting market
structure is an oligopoly. Depending on the kind of
products offered and the way an equilibrium is
obtained, price or quantity setting oligopolies may
be distinguished.
- Product Relationship. Products offered may be
substitutes or complements. In general, we would
expect that products of competing firms are
substitutes. Product differentiation, however,
allows to vary the degree of possible substitution.
- Distribution of Knowledge and Ability. The
distribution of knowledge and ability is closely
related to the possibility of generating sustaining
competitive advantages (Choudhury, Sampler;
1997). If a firm has specific knowledge or specific
abilities that competitors do not have it may use
these skills to outperform competing firms.
- Kind and Degree of Uncertainty Faced by
Competing Firms. Basically we may distinguish
uncertainty with respect to common or private
variables. As an example consider demand
parameters. They are called common or public
variables since they directly affect profits but are
not firm specific. On the other hand variable costs
are an example of private variables (Jin; 1994, p.
323). They are firm specific. Of course, knowledge
of rival’s variable costs may affect a firms own
decisions since it may predict rival’s behavior
more precisely.
- Risk Preferences of Competing Firms. It is
assumed that decision makers are risk averse.
Hence they will not maximize expected profits as
if they were risk neutral but expected utility of
profits.
The results obtained depend on the assumptions
made about the factors identified above. They
partially differ or even contradict each other.
The analysis presented in this paper is carried on
with the help of a microeconomic model based on
game theory. The basic assumption is that
collaboration occurs through knowledge and
information sharing, common information collection
and/or interpretation. In order to share information,
knowledge and know-how, collaborative technology
is usually applied. Joint application development and
joint use of resulting information systems, as well as
inter-organizational information systems in general
are typically covered by such an analysis. Joint
application development bundles development
capabilities in an effort to reduce development costs.
Typically specific know how and information is
shared between the cooperating development
partners. Hence, in case of competing developers, it
is necessary to compare the benefits associated with
reduced costs to possible disadvantages faced by
disclosing information and know how. In this paper
we will assume that information is shared via
temporary joint. Note, that in our context
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
388
collaboration may be characterized as being pre-
competitive. It should not be mixed up with
collusion which may be legally restricted or even
forbidden. A formal model will be developed in the
sequel. Techniques from game theory allow to solve
the corresponding optimization problems. The
model will be analyzed in a simple setting in order
to be able to derive closed-form solutions which
may be handled more conveniently. Hence, it seems
natural to assume that a rational company maximizes
its expected profits. The compensation of such
managers is very often tied to profits. This fact, as
well as possible opportunistic behavior and
asymmetric information, suggest that managers
behave more or less risk averse (Kao and Hughes,
1993). Consequently, expected utility of profits is
maximized instead of expected profits.
3 REINFORCEMENT LEARNING
Learning from reinforcements has received
substantial attention as a mechanism for robots and
other computer systems to learn tasks without
external supervision. The agent typically receives a
positive payoff from the environment after it
achieves a particular goal, or, even simpler, when a
performed action gives good results. In the same
way, it receives a negative (or null) payoff when the
action (or set of actions) performed brings to a
failure. By performing many actions overtime (trial
and error technique), the agents can compute the
expected values (EV) for each action. According to
Sutton and Barto (1998) this paradigm turns values
into behavioral patterns. Most RL algorithms are
about coordination in multi agents systems, defined
as the ability of two or more agents to jointly reach a
consensus over which actions to perform in an
environment. An algorithm derived from Q-
Learning (Watkins, 1989) can be used. The EV for
an action is updated every time an action is
performed, according to Kapetanakis et al (2004):
(1)
Where 01 is the learning rate and p is a
payoff received every time action a is performed.
This is particularly suitable for simulating multi
stage games (Fudenberg and Levine 1998), in which
agents must coordinate to get the highest possible
aggregate payoff. Given a scenario with two agents
(A and B), each of them endowed with two possible
actions
,
and
,
respectively, the agents will
get a payoff, based on a payoff matrix, according to
the combination of performed actions.
3.1 Ego Biased Learning
While discussing the cognitive link among
preferences and choices is definitely beyond the
purpose of this work, it’s important to notice that it’s
commonly accepted that the mentioned aspects are
strictly linked among them. The link is actually bi-
directional (Chen, 2008), meaning that human
preferences influence choices, but in turn the
performed actions (consequent to choices) can
change original preferences. As stated in Sharot et
al. (2009): “…past preferences and present choices
determine attitudes of preferring things and making
decisions in the future about such pleasurable things
as cars, expensive gifts, and vacation spots”.
Even if preferences can be modified according to
the outcome of past actions (and this is well
represented by the RL algorithms described before),
humans can keep an emotional part driving them to
prefer a certain action over another one, even when
the latter has proven better than the former. Some of
these can be simply wired into the DNA, or could
have formed in many years and thus being hardly
modifiable. A bias is defined as “a particular
tendency or inclination, esp. one that prevents
unprejudiced consideration of a question; prejudice”
(www.dictionary.com).
That’s the point behind learning: human aren’t
machines, able to analytically evaluate all the
aspects of a problem and, above all, the payoff
deriving from an action is filtered by their own
perception bias. There’s more than just a self-
updating function for evaluating actions and in the
following a formal reinforcement learning method is
presented which keeps into consideration a possible
bias towards a particular action, which, to some
extents, make it preferable to another one that has
analytically proven better through the trial and error
period. Ego Biased Learning allows to keep this
personal factor into consideration, when applying a
RL paradigm to agents.
In the presented formulation, a dualistic action
selection is considered, i.e.:
,
. In our
context, alternative actions represent two dyadic
collaborative behaviour: the whole of strategic and
tactic enterprise activities to establish a collaboration
or not
. In the last case
, enterprises
carry on other competitive strategies non based on
collaboration. By applying the formal reinforcement
learning technique described in equation (1) an agent
is able to have the expected value for the action it
performed. Each agent is endowed with the RL
technique. At this point, we can imagine two
different categories of agents (
,
: one biased
towards action
and the other one biased towards
COLLABORATION AMONG COMPETING FIRMS - An Application Model about Decision Making Strategy
389
action
. For each category, a constant is
introduced (0
,
1, defining the
expectation of each action, used to evaluate
and
which is the expected value of the
action, corrected by the bias. For the category of
agents biased towards action
we have that:
:
|
|
|
|
(2)
In this way,
represents the propensity for the
first category of agents towards action
and acts as
a percentage increasing the analytically computed
and decreasing
. At the same way,
represents the propensity for the second category
of agents towards action
and acts on the expected
value of the two possible actions as before. The
constant acts like a “friction” for the EV function;
after calculating the objective
it increments
it of a percentage, if
is the action for which the
agent has a positive bias, or decrements it, if
is
the action for which the agent has a negative bias. In
this way, the agent
will perform action
(instead of
) even if
, as long
as
is not less than
.
If
, by definition, the
performed action will be the favorite one, i.e.: the
one towards which the agent has a positive bias.
4 EXPERIMENTAL RESULTS
Some experiments were performed in order to test
the basic EBL equations introduced in paragraph
3.1. The agents involved in the simulation can
perform two possible actions,
and
. The agents
in the simulation randomly meet at each turn (one to
one) and perform an action according to their EV. A
payoff matrix is used, where p
is the payoff
originated when both agents perform a
(both of
enterprises don’t want establish a collaboration), p
is the payoff given to the agents when one of them
performs a
and the other one performs a
and so on
(only one of them want – and try without success –
establish a collaboration. Usually p
and p
are set
at the same value, for coherency. If corss-strategies
are the same and based on collaboration, pay off
p
is maximum. For each time-step in the
simulation, the number of agents performing a
and
a
are sampled and represented on a graph.
In the first experiment, a small bias towards
action
is introduced for fifty
agents (
= 0.1).
Agents
do not have a bias, but all start playing
action
which is the most favourable one,
according to the payoff matrix; this will be different
in the following experiments, where unbiased agents
will start performing a random action. The results
are quite interesting, and depicted in figure 1.
Figure 1: Experiment 1: biased Vs unbiased agents.
In this example it has been chosen to have all the
rational agents (the straight line) starting from the
favoured action (
) so to show that, even so, it is
enough to have one half of the agents acting not
completely rationally to make the system go towards
the sub-optimal action (
). In fact, even if action
is clearly favored by the payoff matrix (payoff 2 vs
1), after taking an initial lead in agents’ preferences,
all the population moves towards action
. This is
due to the resilience of biased agents in changing
their mind; doing this way, the other 50 non-biased
agents find more and more partners performing
action
, and thus, if they perform
they get a
negative payoff. In this way, in order to gain
something, since they are not biased, they are forced
to move towards the sub-optimal action
, preferred
by the biased agents. In order to give a social
explanation of this, we can think to the fact that
often the wiser persons adapt themselves to the more
obstinate ones, when they necessarily have to deal
with them, even if the outcome is not the optimal
one, just not to lose more. This is particular evident
when the wiser persons are the minority, or, as in
our case, in an equal number.
Till now the advantage of performing join action
over
was evident (payoff 2 vs 1)
but not huge; in the next experiment, a new payoff
matrix is used, in the joint action
is
rewarded 3, instead of 2. The purpose is
investigating how much the previous threshold
would increase under these hypotheses. The
empirical finding is 25/75, and the convergence is
again extremely fast, and much similar to the
previous experiment. Even a bigger advantage for
the optimal action is soon nullified by the presence
of just 25% biased agents, when penalty for
miscoordination exists. This explains why
0
20
40
60
80
100
1 3 5 7 9 11 13 15 17 19 21 23 25
Actiona1 Actiona2
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
390
sometimes suboptimal actions (or non-best products)
become the most common. In real world, marketing
could be able to bias a part of the population, and a
good distribution or other politics for the suboptimal
product/service could act as a penalty for unbiased
players when interacting with biased ones.
5 CONCLUSIONS
While individual preferences are very important as a
bias factor for learning and action selection, when
dealing with social systems, in which many entities
operate at the same time and are usually connected
over a network, other factors should be kept into
consideration, when dealing with learning. Ego
biased learning is formally presented in the most
simple case, in which only two categories of agents
are involved, and only two actions are possible
(collaboration or not). That’s to show the basic
equations and explore the results, when varying the
parameters.
Some simulations are run, and the results are
studied, showing how, even a small part of the
population, with a negligible bias towards a
particular action, can affect the convergence of the
whole population. In particular, if miscoordination is
punished (when cross-strategies are different), after
few steps all the agents converge on the suboptimal
action, which is the one preferred by the biased
agents. With no penalty for miscoordination things
are less radical, but once again many non-biased
agents (even if not all of them) converge to the
suboptimal action (non collaborative actions). This
shows how personal biases are important in social
systems, where agents must coordinate or interact.
If we look at things from a managerial/sociologic
point of view, we have the following explanation.
The presented experiments show that few players
potentially adverse to exchange information in a
system, are enough for all the players to stop
exchanging. This happens because the higher risk
aversion of these operators brings all the others to
the idea that carrying on collaborative strategies is a
potential dispersion of resources. In fact, whenever a
collaborative player crosses a non-collaborative one,
they both evaluate the possible business, but after
that the non-collaborative player denies it. From this,
the penalty for miscoordination. A collaborative
rational agent, after meeting some non-collaborative
ones, changes her mind as well, since each time she
loses some resources. Then, she becomes non-
collaborative as well, unless she finds many
collaborative players in a row. In other terms, to
avoid a refusal after trying to collaborate, which is
something that waste time and resources, also
potentially collaborative agents will start to
immediately refuse the possibility of a cooperation.
By doing this, they won't gain as much as they
would through collaboration, but they won't also risk
to lose resources. The whole system thus settles on
the sub-optimal equilibrium, in which no player
collaborates.
In future works, general cases will be faced
(more than two possible actions, different biases) in
order to analyze the psychological drivers behind
firms collaborations and additional experiments will
be run.
REFERENCES
Chen M. K., 2008. Rationalization and Cognitive Disso-
nance: do Choices Affect or Reflect Preferences?
Cowles Foundation Discussion Paper No. 1669
Clemons, E., Reddi, S. Row, 1993, M.: The Impact of
Information Technology on the Organization of
Economic Activity – The "Move to the Middle"
Hypothesis. Journal of Management Information
Systems 10, No. 2, pp. 9-35.
Fudenberg, D., and Levine, D. K. 1998. The Theory of
Learning in Games. Cambridge, MA: MIT Press
Jin, J.: 1994. Information Sharing Through Sales Report.
Journal of Industrial Economics 42, No. 3
Kao, J., Hughes, J. 1993. Note on Risk Aversion and
Sharing of Firm-Specific Information in Duopolies,
Journal of Industrial Economics 41, No. 1
Palfrey, T. R., 1982. Risk Advantages and Information
Acquisition. Bell Journal of Economics 13, No. 1, pp.
219-224.
Powers R. and Shoham Y., 2005. New criteria and a new
algorithm for learning in multi-agent systems. In Pro-
ceedings of NIPS.
Sharot T., De Martino B., Dolan R.J., 2009. How Choice
Reveals and Shapes Expected Hedonic Outcome. The
Journal of Neuroscience, 29(12):3760-3765
Sutton, R. S. and Barto A. G., 1998. Reinforcement Learn-
ing: An Introduction. MIT Press, Cambridge, MA. A
Bradford Book
Watkins, C. J. C. H., 1989. Learning from delayed
rewards. PhD thesis, Psychology Department, Univ. of
Cambridge.
Williamson, O., 1975. Markets and Hierarchies – Analysis
and Antitrust Implications. New York.
COLLABORATION AMONG COMPETING FIRMS - An Application Model about Decision Making Strategy
391
