Auction Model of P2P Interaction in Multi-Agent Software
Anton Ivaschenko
1
and Andrey Lednev
2
1
Magenta Technology, 68, Lombard Street, City of London, EC3V 9LJ, London, U.K.
2
Samara State Technical University, Molodogvardeiskaya street, 244, Samara, Russian Federation
Keywords: Integrated Information Space, P2P, Auction, Interaction Management.
Abstract: The paper describes one of the possible models of interaction management of active software agents in P2P
network of the enterprise information space. The matrix form of enterprise management is being projected
on the P2P network of interacting software components. It is proposed to study the problem of resource
allocation using Auction model enhanced by the opportunity of its management by changing dynamical
characteristics. The approach shows that the introduction of delays and accelerations allows to increase the
efficiency of solving scheduling problems.
1 INTRODUCTION
Modern trends in the area of enterprise management
consider various organizational models. For
example, a matrix model is actively used nowadays
to represent the processes of collaborative decision
making according to coordinated staff interaction.
Integrated information space is required to support
such structures in order to automate decision making
processes.
Network architectures of bundled interacting
software solutions should be created to form such an
environment. The concept of building a distributed
heterogeneous integrated information space with
P2P networks (Schoder and Fischbach, 2003) of
software agents capable of interaction by means of
messages’ exchange (Lednev, 2010) becomes
widely spread.
The principles of bio-inspiration (Leitao, 2009)
are theoretically used for solving design problems of
such information environments that allow
considering users and software agents as a complex
evolving system. In practice this concept is being
successfully applied in implementation of multi-
agent technologies (Andreev and Glashchenko,
2009).
However the questions of organizing the
management in such systems are still open. The
reasons of it are: an impossibility to apply direct
instructions, an infeasibility of classical optimization
and a whole complexity of virtual multi-agent world.
One of the most perspective approaches to deal
with these problems is to provide indirect
informational management enhanced by methods
and means that are focused on a rhythmicity analysis
of messages’ exchange between the interacting
participants: enterprise employees and software
agents as their representatives.
In this paper we present an opinion on how to
organize such an interaction in a form of auction.
Such approach can be helpful for development of
intellectual automated resources scheduling systems
in different problem domains, for example in
transportation logistics.
2 AUCTION BASED
INTERACTION IN
INTELLIGENT SCHEDULING
One of the most widespread areas of multi-agent
technology application is real time scheduling of
mobile resources in transportation logistics. Real-
time mode means an ability to incrementally
schedule the orders that arrive at various times
providing the minimal deviation from delivery time
with the maximal load of transport (to reduce
backhaul run).
Driver involvement in decision making will be
very helpful for finding appropriate solutions as it
will allow considering specific order requirements.
In response to such individual approach dispatcher
expects the growth of overall service quality. Thus
431
Ivaschenko A. and Lednev A..
Auction Model of P2P Interaction in Multi-Agent Software.
DOI: 10.5220/0004329904310434
In Proceedings of the 5th International Conference on Agents and Artificial Intelligence (ICAART-2013), pages 431-434
ISBN: 978-989-8565-38-9
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
dispatcher is “selling” orders to drivers inspiring
them to provide better service.
Described interaction can be technically
organized with modern info communicational
technologies. In order to automate the solution when
driver requirements are known preliminary a
negotiation process can be simulated in the virtual
environment organized in the form of P2P network
of software agents that represent decision-makers.
Resource distribution in the described environment
can be organized on the basis of auction model.
According to this model the order can be treated as a
lot, dispatcher as an auctioneer and drivers as
bidders. The strategic aim of this auction is to
maximize the load of resources that increases the
risks of fail to serve the order and make the whole
schedule inconsistent.
Auction can be defined as a public sell of one lot
according to the predefined rules. The auctioneer or
the Centre (the agent that represents a centralized
dispatcher) at different moments of time exposes lots
(orders), which are of different level of interest to
the bidders (drivers or their agents). The winner of
the auction is an agent who buys a lot according to
the rules. The aim of an agent is to win as much lots
as possible that are of a particular interest for him at
minimal prices.
The proposed auction is divided into an uncertain
number of iterations. Blind bids are made during the
iterations, and the Centre announces the highest bid
afterwards. The Centre announces the price for new
iteration and the approximate duration of the
following iteration. In response any of the
participants can increase the bid and thus initiate a
new iteration. Auction is finished when no new bids
are made during the iteration.
In real auction auctioneer makes pauses between
gavel heats to stimulate making higher bids faster.
The same way in multi-agent information space the
Centre activates interaction between software
components by setting up the timeframes of
iterations. Therefore the Centre provides indirect
information management.
So the following 2 ideas can be proposed to
provide information management of multi-agent
intelligent scheduling:
1) implement the auction model for organization
of effective resource allocation in P2P distributed
information space;
2) manage agents’ negotiations in a process of
auction based resources allocation by varying time
intervals of the bidding iterations.
The difference of P2P auction from an ordinary
multi-agent auction is that the Center as P2P node
addresses each bidder individually that allows
implementing a method of agent’s adaptive
management by information. This method is based
on adaptive limiting, scoping and garbling being
applied to each node individually. For example, a
taxi dispatcher can inform different drivers only
about filtered orders to provide better cabs
geographical distribution and higher service level.
3 AN AUCTION MODEL FOR
AGENTS’ INTERACTION
In the following section the formal model of the
proposed approach is presented. For this purpose the
following variables are introduced:
0
C
is an initial price of a lot;
Niv
i
...1,
is a value of a lot for every particular
agent (N is agents’ total number);
ji
c
,
is a price of the bid made by an actor;
Mj ...1
is a number of iteration;
j
t
is an iteration’s duration.
The durations of iterations are not equal, as well as
not fixed, since they are started by new price
announcements made by the Center. New lot price
ji
C
,
is chosen as a maximum between all bids.
Auction is finished after the following time
passes since its start:
M
j
jA
tT
1
(1)
At each iteration the Centre is interacting with
agents using P2P principle: every message contains
a new bid proposal of lot price at the moment
ji
t
,
:
jijiji
tCs
,,,
,
(2)
Response messages can be sent by agents in turn:
jijiji
tcb
,,,
',
(3)
where
jijiji
cCc
,,,
.
Time required for an agent to make a decision is
,,
'
ij ij
tt
.
In such a model some agents can miss iteration and
make no bids. The goal of the Centre that organizes
Kk ..1
auctions can be defined as:
minmax,max
1
,
1
,
K
k
kA
K
k
Mi
i
TC
(4)
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
432
The goal of an agent is the following:
minmax
max,max
1
,,,,,,
1
,,,,,
K
k
kMi
i
kMikMi
K
k
kMi
i
kMiki
Ccc
Ccv
(5)
where
.0,1
;0,0
x
x
x
is step function.
This means that an agent can be satisfied by buying
lots at minimal price, and it doesn’t matter what time
it takes. The Centre to achieve its own goal in turn
can change the durations of iterations and the
number of actors involved (deciding whom and
when to send the messages
ji
s
,
).
In order to sell a lot at its highest price the Centre
needs to organize a competition between the agents,
which requires a proposal request strategy (plan).
Every agent can manage the size and time of its own
bid thus arising interest to itself from the Centre. On
contrary to achieve equilibrium the Centre should
announce its plan to the agents. In this case it is
possible to study the dependence between the plan
of the Centre and the strategies of agents.
As an example of the proposed interaction model
implementation in practice the following resources
allocation problem can be considered. Nowadays
many shops practice separation of sales to several
stages. At every next stage the price of the good is
decreased. Thus shop tries to get rid of out-of-date
things and renew the assortment getting as max
profit as possible.
On the other hand buyers want to buy the goods
they need at the lowest price. Still a customer can’t
be confident what time to wait for a sell-out:
somebody else can be ahead or the shop can stop
“the game”. A shop can also fix the price thus
making the wait of a customer senseless. This can be
simulated by a model of simultaneous numerous
auctions in several iterations with a single Centre
and unknown number of agents.
The problem statement can be simplified by
depriving the agents to affect the price of a lot. In
this case the time factor becomes determinative. It
should be mentioned that constant price change of
the bid can’t affect the resulting lot price. In case
when actors represent no interest (make no bids) the
Centre is forced to lower prices faster.
4 SIMULATION
To study the proposed approach some simulation
modeling was performed aimed to identify the
opportunity of increasing resulting lot price by
managing iteration time. Simulation included two
experiments in real time with different durability of
lot sale. In order to carry out the experiments a
special software environment was implemented in
J2EE modeling auction based interaction of agents
in distributed information space.
The Centre specified the time after which the
agents had to make bids. Among the proposed bids
the highest one was chosen as the price of the lot for
the next iteration. The size of the bid made by an
agent was calculated randomly using normal
distribution against the current price of the lot.
Neither the Centre nor the agents knew the time
when auction would be finished.
Figures 1 – 3 illustrate the results for 2 agents’
auction. In the first case (without time management)
iteration’s duration was constant: 41% of lots were
sold within 3 iterations and other 59% within 4. In
the second case (see Figure 2) The Centre was
enabled to manage iterations duration (see Figure 3).
Figure 1: Bids during iterations without time management.
AuctionModelofP2PInteractioninMulti-AgentSoftware
433
Figure 2: Bids during iterations with time management.
Figure 3: Number of iterations in experiments with time management.
So it reduced the time of iterations proportionally
the sizes of the bids: the higher was the bid the
shorter was the following iteration. As the result
approximately from 5 to 15 iterations were made
while the lot was played.
The difference between lot prices appear in both
cases, which can be explained by the normal
distribution of bid prices. However in general the
simulation results proved the idea that the final price
of the lot is be higher if it is played within more
iterations.
5 CONCLUSIONS
The proposed approach and a model for auction
based interaction of agents in integrated information
space allow implementation of multi-agent software
for P2P networks management. They expound the
ideas of indirect management by information that
corresponds with the concepts of bio-inspired
approach and can be extended by the methods of
conditional management.
The results of simple simulation prove the
possibility to use auction models in solving practical
problems of intelligent scheduling, for example in
transportation logistics, and can be helpful for
automated decision making support in real time.
REFERENCES
Schoder, D., Fischbach, K., 2003. Peer-to-peer prospects.
Communications of the ACM, vol. 46, no. 2. p. 27-29
Lednev, A., 2010. Mobile P2P taxi service.
MSc
Dissertation
, University of Surrey. 75 p.
Leitao, P., 2009. Holonic rationale and self-organization
on design of complex evolvable systems.
HoloMAS
2009, LNAI 5696, Springer-Verlag Berlin Heidelberg.
p. 1-12
Andreev, V., Glashchenko, A., Ivaschenko, A.,
Inozemtsev, S., Rzevski, G., Skobelev, P., Shveykin,
P., 2009. Magenta multi-agent systems for dynamic
scheduling.
ICAART 2009 International Conference
on Agents and Artificial Intelligence
. p. 489-496
ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence
434
