Information Technology Reference
In-Depth Information
Monotonic Mixing of Decision Strategies for
Agent-Based Bargaining
Jan Richter, Matthias Klusch, and Ryszard Kowalczyk
Swinburne University of Technology, Melbourne, Australia
German Research Center for Artificial Intelligence, Saarbrucken, Germany
{
jrichter,rkowalczyk
}
@swin.edu.au, klusch@dfki.de
Abstract.
In automated bargaining a common method to obtain
complex concession behaviour is to mix individual tactics, or decision
functions, by a linear weighted combination. In such systems, the ne-
gotiation process between agents using mixed strategies with imitative
and non-imitative tactics is highly dynamic, and non-monotonicity in
the sequence of utilities of proposed offers can emerge at any time even
in cases of individual cooperative behaviour and static strategy settings
of both agents. This can result in a number of undesirable effects, such
as delayed agreements, significant variation of outcomes with lower util-
ities, or a partial loss of control over the strategy settings. We propose
two alternatives of mixing to avoid these problems, one based on individ-
ual imitative negotiation threads and one based on single concessions of
each tactic involved. We prove that both produce monotonic sequences
of utilities over time for mixed multi-tactic strategies with static and dy-
namically changing weights thereby avoiding such dynamic effects, and
show with a comparative evaluation that they can provide utility gains
for each agent in many multi-issue negotiation scenarios.
1
Introduction
Automated negotiation between rational software agents is considered key to
facilitate intelligent decision-making between two or more parties which are in
conflict about their goals. In such environments, the agents acting on behalf of
their users (buyers, sellers) have no or only uncertain knowledge about oppo-
nent's behaviours and can use a range of different strategies to conduct nego-
tiation. In automated bargaining or bilateral negotiation, two rational agents
negotiate by alternatively exchanging offers over issues of a service or product
where each agent has the preference to achieve the highest possible utility from
an outcome while the common interest is to find an agreement before the dead-
line. A common approach for the agents to propose offers is to use individual
decision functions, also called tactics, and mix them based on linear weighted
combinations to create complex concession behaviour in the form of negotiation
strategies. For instance, the prominent service-oriented negotiation model by
Faratin et al [4] proposes different types of tactics such as behaviour-, time- or
resource-dependent that can be mixed together. In this paper, we demonstrate
