COGNITIVE MODELING OF INTERACTIONS

DURING A NEGOTIATION PROCESS

Charlène Floch, Nathalie Chaignaud and Alexandre Pauchet

LITIS - EA 4108 - Place Emile Blondel - BP 08 - 76131 Mont-Saint-Aignan Cedex, France

Keywords: Cognitive Modeling, Negotiation, Interaction, Multi-Agent Systems.

Abstract: This article presents a study on human negotiations in order to improve the BDI

GGY model of agent with

negotiation skills. This study is based on logs of real negotiation rounds, obtained through a psychological

experiment. We propose an utterance model including performatives applied to mental states and a dialog

model using timed automata, both based on the BDI concepts.

1 INTRODUCTION

Negotiation is required in various situations: by

phone in a trading room, in an interview with one’s

manager to discuss a raise. In the context of e-

business, negotiation can be required between

humans and/or artificial agents. Be it in a

commercial, social or political setting, it is above all

a human process. In the animal kingdom, the

balance of power removes all possible negotiations.

In order to enable negotiation between artificial

agents and humans, it is necessary to first study the

processes set up between humans. Therefore, we aim

here at modeling human negotiation between a seller

and a buyer. The originality of this work stems from

the fact that it is based on a psychological

experiment where subjects are involved in a

negotiation task. We are interested in the

formalization of the messages sent and their

dynamics, without taking into account the decision

making of the subjects in itself.

To extend the model of agent called BDI

GGY

(Pauchet et al., 2007) by adding negotiation skills,

we derive negotiation protocols for software agents

from the observation of human negotiation during

the experiment. Section 2 describes related work on

negotiation and Section 3 presents the BDI

GGY

system. The psychological experiment is explained

in Section 4 and the analysis of the logs is given in

Section 5. Section 6 proposes a model of interaction

for negotiation. In Section 7, conclusion and

perspectives close this paper.

2 NEGOTIATION IN MAS

Negotiation is addressed here in the point view of

multi-agent systems (MAS).

2.1 Notion of Negotiation

(Rahwan et al., 2007) propose the following

definition of negotiation: “a form of interaction in

which a group of agents, with conflicting interests

and a desire to cooperate, try to reach a mutually

acceptable agreement on the division of scarce

resources”. However, negotiation does not concern

only resource sharing. We can also negotiate

interpretations, responsibilities, …

(Chang and Woo, 1994) emphasize the

competition aspect of negotiation. Negotiation

begins with the observation of a disagreement

caused by dissimilarities of preferences. This

definition is restrictive because negotiation can fail

to reach a compromise.

Considering that these definitions are not

satisfactory enough, we propose the following

formulation: “the negotiation is the process to find

an agreement between several players having

different preferences that may change with time”.

2.2 Models of Negotiation

The literature on models of negotiation between

software agents or humans is extensive.

Some researchers are interested in modeling

preferences and in the mechanisms of negotiation

221

Floch C., Chaignaud N. and Pauchet A. (2008).

COGNITIVE MODELING OF INTERACTIONS DURING A NEGOTIATION PROCESS.

In Proceedings of the Tenth International Conference on Enterprise Information Systems - HCI, pages 221-226

DOI: 10.5220/0001698202210226

 SciTePress

(Bellosta et al., 2004). Others develop applications

in e-business (Chavez and Maes, 1996) and solve

resource allocation problems with utility functions.

Negotiation by argumentation is treated in (Kraus et

al., 1998) and (Schroeder, 1999). (Bartolini et al.,

2002) and (Matthieu and Verrons, 2004) develop

generic systems by implementing different types of

negotiation.

2.3 Negotiation Protocols

The first example of protocol is Contract Net (Smith,

1980), which makes a manager and contractors

interact through contract exchanges. Contractors are

responsible for the task execution and for the result

transmission, whereas the manager is responsible for

the task managing and for the result treatment.

The Sian protocol (Sian, 1991) allows a MAS to

interpret objects of the environment. The perception

is distributed and the agents negotiate in order to

choose a common interpretation of the environment.

The agents interact via a blackboard. The discussion

always finishes with an agreement consisting in

accepting or withdrawing a hypothesis on the

blackboard.

The SANP protocol (Chang and Woo, 1992)

models a negotiation between two agents. The

speaker tries to make his proposition accepted by the

interlocutor. The protocol comprises several stages:

the beginning, the bid formulation and the receipt of

the answer (acceptance or refusal), the attack, the

tactic, the problem solving and the result. According

to the reasoning of the agent, a set of strategies and

automata corresponds to each phase. To evaluate

their model, Chang and Woo asked subjects to use

their system to solve conflicts. Their method is very

different from ours because they get human involved

once the model is built. In our approach, we obtain

behaviors of negotiation by experiment and the

model is built according to what we observed.

3 THE BDIGGY SYSTEM

The BDIGGY system (Pauchet et al., 2007) was built

to understand and simulate how human elaborate

plans in situations where knowledge is incomplete

and how they interact to obtain missing information.

It implements a human planning model and a human

interaction model through the BDI concept. We

detail here only the interaction model.

3.1 The Dialog Model

To consider sequentiality and temporality of the sent

messages, exchanges are represented through timed

automata. Managing time is useful when subjects

tend to delay their answer to a question. This can

lead to sending re-queries and the exchange can be

considered closed by the protagonists without an

explicit ending message.

Eight (4x2) automata have been built: one for

each interlocutor and one for each type of exchange

(information query, information proposal,

spontaneous sending, error processing).

3.2 The Utterance Model

To refer to the speech act theory, the observed

speech acts (performatives) are either descriptives,

directives or commissives. A performative is applied

to a mental state (a belief or a desire) and is linked to

the content of the message:

• A descriptive is applied to a belief. It describes

how the sender perceives the world.

• A directive is applied to a desire of the sender.

The sender wants to receive information and he

sends this desire to the subject who has the

information.

• A commissive is applied to a desire of the

receiver. The sender supposes that the receiver

has a certain desire.

The semantics of each performative is given by a

generic reduction rule describing formally the pre-

conditions of the sending and the reception of a

message and the actions to be done. It is linked with

the timed automata and the mental states of the

modeled agent.

3.3 The BDIGGY Architecture

BDIGGY is based on the BDI concept and includes:

• a perception module, which analyzes the

environment and generates beliefs,

• a human planning module, which builds abstract

plans in an opportunistic way,

• a plan interpreter, which works as the BDI

interpreter,

• and a communication module, which

implements the human interaction models (the

utterance model, the dialog model, the

semantics of the performatives).

ICEIS 2008 - International Conference on Enterprise Information Systems

222

4 THE EXPERIMENTAL

FRAMEWORK

This section describes the experiment we have set up

to observe human behaviors and record logs during

interactions of negotiation.

4.1 The Problem

The problem we chose is the game « Négoces »

(http://www.negoces.fr), in which the players are

compelled to negotiate to reach their goals.

At the beginning of the game, each player

receives fifteen gold coins (that all players can see)

and eight “merchandise” cards (that the other players

cannot see) comprising two series of four identical

cards (corresponding to exclusive control on two

merchandises). To win, a player has to obtain eight

different cards, keeping at least ten gold coins. Time

is not limited. There are several auctions during one

game.

The first round enables to select the initial seller.

The card “Diamond” is sold by auction but there is

no negotiation. All the subsequent auctions are done

by negotiation. The player who has nine cards is the

seller. He chooses one of his cards and proposes it

for auction. The buyers negotiate with the seller the

acquisition of the card, in any order. They can

propose one card, gold coins, or one card plus gold

coins. One offer can increase, decrease or be

removed until the auction of the card. The buyers

cannot exchange cards each other. The seller will

eventually have to accept an offer, even if he is not

satisfied.

The discussion is limited to the current auction

until the transaction is done (exchange of cards and

coins). One game comprises several auctions (at

least five) until one player has won.

4.2 The Experiment

In order to observe the behaviors of the players, we

developed a system allowing to play this game over

the network. The subjects are isolated in different

rooms and they interact through the user-interface. It

contains a communication panel, allowing the

players to chat, and an action panel, to perform

actions related to one auction. The possible actions

are different according to the player (seller or

buyer): the seller can sell a card and accept an offer,

whereas the buyers can propose an offer or leave.

All actions and all interactions during the game have

been recorded in a text file called log.

We chose to imply four players. This number is

sufficiently high to expect complex negotiation

behaviors like a coalition and sufficiently low to

carry out the experiment. Fourteen groups of four

players (secretary, PhD students, researchers and

teachers) participated to our experiment. We had to

remove one of these groups because the subjects had

difficulties to understand the rules of the game.

Thus, we obtained thirteen usable logs.

A fragment of log is presented below:

[15:28:10] 127, ACCEPT_SELL (Jack, COINS -, CARD

honey, Averell, COINS -, CARD coffee)

[15:28:33] 128, William SAY I have to be careful!

[15:28:41] 129, Jack SAY Who does want a card

cigar or a card spice?

[15:28:58] 130, Averell SAY I propose 2 coins

[15:29:16] 132, Jack SAY Averell, I am ok

[15:29:27] 133, Jack AUCTION cigar

[15:29:29] 134, William SAY I propose 3 coins for

cigar

[15:29:34] 135, William PROPOSES (Jack, COINS -,

CARD cigar, William, COINS 3, CARD -)

[15:29:35] 136, Averell TAKE_OUT

[15:29:36] 137, Jack SAY Ok William

[15:29:37] 138, ACCEPT_SELL (Jack, COINS -, CARD

cigar, William, COINS 3, CARD -)

This log corresponds to only one auction.

Players are named Averell, William, Joe and Jack to

preserve anonymity. The lines with the word “SAY”

correspond to interactions in the chat panel and

others correspond to actions.

5 ANALYSIS OF LOGS

This section presents our analysis of the thirteen

logs.

Although the situation implies four players,

negotiations are always established between the

seller and one buyer. The buyers rarely talk to each

other. But when they do, they use humor or they

comment on what the others do and we decided not

to model these interactions.

One auction can be broken down into exchanges

according to the discursive goal of a player. These

exchanges are classified into the following types:

• Auction (from the seller),

• Information query,

• Reminder,

• Discussion (to achieve an agreement),

• Spontaneous sending of information,

• Warning.

All these types of exchanges are not used during

each auction but there is at least one auction and one

discussion. The discussion can begin in different

ways: the buyer proposes an offer to the seller or the

seller asks for an offer from the buyer.

COGNITIVE MODELING OF INTERACTIONS DURING A NEGOTIATION PROCESS

223

In the second case, the seller actually wants to

negotiate with the buyer. This is more than a

reminder: the seller initiates the negotiation.

The negotiation process does not always begin

with the auction of one card. That is why the auction

is differentiated from the discussion between the

protagonists. In six groups, more than 50% of the

sells begin with an action different from

auction.

The auction begins with a discussion about the card

to be sold. The players send information queries,

offers and reminders.

The discussion between the seller and a buyer

ends either when the buyer leaves, or when the seller

accepts or refuses the offer.

During the information queries, many questions

remain unanswered. It seems that the players are

intensely focused on the game and that the speed of

offers and interactions is too high. In some cases, the

players do not want to answer (for example, to

conceal their owned cards).

As for the reminders, they can arise from the

seller or a buyer.

We also observed menaces, coalitions, promises

and conditional offers. But these interactions were

not numerous enough to be taken into account in our

study.

Table 1: List of observed performatives.

Performative

Description

Auction The seller sells one card

Joe SAY I sell tea

takeOut

One buyer leaves

Jack SAY I leave

inform A player sends information spontaneously

Jack SAY I am interested in spices

query / reply With a query, a player asks information to an

other player who answers with a reply

Joe SAY who can offer to me a card honey

against a card cigar?

Jack SAY I do.

bid

A buyer makes an offer

William SAY I offer 5 coins and tee

acceptBid

The seller accepts an offer with acceptBid.

Then, he informs the others with sold

Jack SAY 4 coins, ok !

refuseBid A player refuses an offer

Jack SAY 1 coin is not enough

request

The seller asks the buyers to make new offers,

or to modify an offer

Jack SAY push up Joe

relaunch

A player reminds other players

Averell SAY, Go on Jack!

warning

A player warns other players about an offer

Averell SAY If you sell to Joe, he will win!

All the logs were manually analyzed. The list of

the speech acts observed in the logs is given in

Table 1. Three of them (inform, query and reply) are

already used in the BDI

GGY model.

Table 2 presents the observed types of

exchanges. Each of these exchanges is guided by the

dialogical goal of the speaker, according to the first

performative he sent.

Table 2: List of observed types of exchange.

Types of

exchange

Dialog

goal

First

performative

Closing

performative

information

query

directive query reply / -

commissive

of the buyer

bid

acceptBid /

leave

discussion

directive

of the seller

request acceptBid /

leave

spontaneous

sending

descriptive inform inform

reminder directive relaunch relaunch

warning descriptive warning warning

Each exchange has been annotated in the logs by

its first performative and its type.

None of the models mentioned in Section 2

enable to represent entirely these observed

exchanges. For example, they do not allow

sequences such as refuse/accept or modify/accept:

the seller can accept an offer he has previously

refused or an offer he asked for modification without

success. Decisions made by the players can evolve

during the negotiation process according to the

interactions and time that goes by.

The speech act list should not constrain the

interactions between the protagonists.

Moreover, according to us, a contract is the result

of a negotiation and its modalities are discussed.

However, in the Contract Net protocol, the manager

proposes the contract and the contractors cannot ask

for modification of it. We think that the contractors

should have the possibility to modify the current

contract and they do not have to create a new one.

Among our needs, the following possibilities are

not offered by other models:

• give an answer between agreement and refusal,

• propose modifications for all protagonists ,

• express the reason of a refusal,

• decide not to enter the negotiation…

It is more the question of negotiating the content

of a contract than the question of signing an existing

contract.

ICEIS 2008 - International Conference on Enterprise Information Systems

224

6 THE INTERACTION MODEL

This section describes our interaction model: the

dialog model and the utterance model.

6.1 The Dialog Model

As mentioned in Section 5, the players can decide

not to answer a question, or refuse to negotiate or

leave the negotiation. We have to model these

behaviors to make software agents as human-like as

possible. BDI

GGY already models a lack of response

through a timed automaton. This model has to be

extended in order to take into account the

negotiation processes observed in the logs.

The negotiations are established between a buyer

and the seller. Thus, the automata can be opened

between two players. Our model includes ten

automata (5x2): one for the speaker and one for the

interlocutor of each type of exchange.

The set of timed automata used is: Q

ini

, Q

int

, I

ini

int

, D

ini

, D

int

, W

ini

, W

int

, R

ini

, R

int

, A

ini

, A

int

, where

{Q, I, D, W, R, A} is the type of exchange (Q:

information query, I: spontaneous sending, D:

discussion, W: warning, R: reminder, A: auction)

and {ini, int} describes the protagonists of the

exchange (ini for the speaker and int for the

interlocutor).

6.1.1 Automata Coming from BDIGGY

Two automata present in the BDIGGY model are

reused for the negotiation: the information query

automaton and the spontaneous sending automaton.

Only the former is described here (Figure 1): it is

only partly reused.

Figure 1: Information query automaton (speaker).

The transitions observed in the protocols of our

application are colored in red.

6.1.2 Discussion Automata

As already said, the negotiation discussion begins

either when the buyer makes an offer or when the

seller asks for an offer from the buyer. Thus, the

discussion exchanges were broken down into two

types (Buyer or Seller) and in fact into four

automata: D

ini

Buyer, D

int

Seller, D

ini

Seller and

int

Buyer. Figure 2 shows the D

ini

Buyer automaton

used when the buyer initiates the discussion (bid).

The seller opens the symmetric automaton D

int

Seller.

The similar automaton D

ini

Seller corresponds to

the fact that the seller asks for an offer from the

buyer. It differs only from the first performative sent

(request). As for the buyer, he opens the D

int

Buyer

automaton.

6.2 Utterance Model

In our application, negotiations are applied to the

auction of one card. It consists on specifying the

seller, the buyer, the card and the number of coins

exchanged.

pTrans is a predicate that represents an

auction:

pTrans(a1, c1, tc1, a2, c2, tc2)

where

a1 is the seller, c1 the number of coins

given by a1, tc1 is the card that a1 sells, a2 is the

buyer,

c2 is the number of coins given by a2, tc2 is

the card given by a2.

A message exchanged by the agents is

represented by the predicate

pMessage(A

P O)

where A

is the sender, A

is the receiver, P is the

performative used,

O is the object on which the

performative is applied.

As for BDI

GGY, each utterance is represented by

a performative applied to a mental state. For

example, all the directives are applied to a desire of

the speaker, their general form is:

pMessage(A

P pD(As,E))

where pD is a desire of the agent A

that is applied

to the predicate

E (pTrans).

Semantics of each performative are given by two

general reduction rules (one for the sending and one

for the receiving), with the pre-conditions of the

sending (respectively receiving), the states of the

automata from which the performative can be sent

(respectively received), the new current state in the

opened automaton and the updates to be done in the

memory of the agent.

COGNITIVE MODELING OF INTERACTIONS DURING A NEGOTIATION PROCESS

225

Figure 2: Discussion automaton of the buyer.

7 CONCLUSIONS AND

PERSPECTIVES

The interaction model for the negotiation proposed

in this article is based on the analysis of

experimental logs. It aims at simulating as faithfully

as possible the human processes of negotiation.

New performatives have been proposed, linked

with timed automata to model the utterance level

and the dialog level of interactions. This model is

adaptable and enables the players to change their

mind according to the situation and time. It also

enables to modify the offers at any time.

However, the reasoning model is not built yet.

We are currently analyzing the experimental

protocols from a reasoning point of view.

The BDI

GGY model is also generic enough to be

re-usable for this new application (negotiation

between a seller and several buyers).

When the reasoning model will be included in

BDI

GGY, the validation of the system will be done

by comparing the set of artificial logs and the set of

human logs (for example, with the use of a Turing-

like test and with the use of hypothesis testing).

REFERENCES

Bartolini C. Preist C, Jennings N, 2002. Architecting for

reuse: A software framework for automated

negotiation, International Workshop on Agent-

Oriented Software Engineering, Bologna, Italy.

Bellosta M-J, Kornman S, Vanderpooten D, 2005. A

framework for multiple criteria English reverse

auctions, IAT’05, Compiègne, France, 633-639.

Chang M, Woo C, 1992. SANP: A Communication Level

Protocol for Negotiations, Decentralized AI, (3), 89-

198.

Chavez A, Maes P, 1996. Kasbah: An Agent MarketPlace

for Buying and Selling Goods, PAAM’96, 75-90.

Kraus S, Sycara K, Evenchik A, 1998. Reaching

agreements through argumentation: a logical model

and implementation, Artificial Intelligence, 104, 1-69.

Mathieu P, Verrons M-H, 2005. A Generic Negotiation

Model using XML, AISB, 1 (6).

Pauchet A, El Fallah-Seghrouchni A, Chaignaud N, 2007.

A Computational Model of Human Interaction and

Planning for Heterogeneous Multi-Agent Systems,

AAMAS'07, Hawaï, 391-393.

Rahwan I, Sonenberg L, Jennings N, McBurney P, 2007.

STRATUM: A Methodology for Designing Heuristic

Agent Negotiation Strategies, Applied Artificial

Intelligence, 21 (10) 41.

Schroeder M, 1999. An Efficient Argumentation

Framework for Negotiating Autonomous Agents,

MAAMAW, Valancia, Spain.

Sian S, 1991. Adaptation based on cooperative learning in

multi-agent systems, Decentralized AI 2, Elsevier

Science, 257-272.

Smith R, 1980. The Contract Net Protocol: High-Level

Communication and Control in a Distributed Problem

Solver. IEEE Transactions on Computers, 29(12),

1104-1113.

ICEIS 2008 - International Conference on Enterprise Information Systems

226