Agent Technology for e-Commerce Chapter 12: Trust, Security and Legal Issues Maria Fasli .
Agent Technology for e-Commerce Chapter 9: Negotiation II Maria Fasli .
-
date post
22-Dec-2015 -
Category
Documents
-
view
220 -
download
4
Transcript of Agent Technology for e-Commerce Chapter 9: Negotiation II Maria Fasli .
Agent Technology for e-Commerce
Chapter 9: Negotiation II
Maria Faslihttp://cswww.essex.ac.uk/staff/mfasli/ATe-Commerce.htm
2Chapter 9
Agent Technology for e-Commerce
Bargaining
A bargaining situation: two or more agents have a common interest and could reach a mutually beneficial agreement, but have a conflict of interest about which one to reach
Seller’s surplus
Buyer’s surplus
Agreement zone
ps pb
£
Seller’s valuation:wants to receiveps or more
Buyer’s valuation:wants to pay pb
or less
Agreement price p*
Buyer wants to decrease p*
Seller wants to increase p*
3Chapter 9
Agent Technology for e-Commerce
Bargaining power
The bargaining power of the participating agents in a bargaining situation is determined by a number of factors
Impatience Risk of breakdown Outside options Inside options Commitment tactics Asymmetric information
4Chapter 9
Agent Technology for e-Commerce
Axiomatic bargaining
Axiomatic bargaining theory assumes no equilibrium Axiomatic models of bargaining yield solutions that satisfy a set
of desired properties – the axioms of the bargaining solution
Example Two agents A and B need to divide a cake of size The set of possible agreements that they can reach is:
={(oA,oB):0oA and oB= -oA
The agents’ utilities are:
UA (oA)= uA and UB (oB)= uA
If the agents fail to reach a deal, then a default solution is implemented and they gain utility (dA, dB)
5Chapter 9
Agent Technology for e-Commerce
Nash Bargaining Solution
The Nash Bargaining Solution (NBS) of this bargaining situation is the allocation of utilities (uA, uB) which solves:
o*=max(uA- dA) (uB- dB) subject to (uA dA) and (uB dB)
The NBS is the only bargaining solution that satisfies the following: Pareto efficiency Symmetry Invariance Independence of irrelevant alternatives
6Chapter 9
Agent Technology for e-Commerce
Returning to the example:
uA= oA and uB= oB = (1- oA ) The NBS is the sharing rule that maximizes the Nash product:
(oA - dA) (oB - dB) The NBS is at:
uA=[+dA-dB]/2 and uB=[+dB-dA]/2
uA=dA +[-dA-dB]/2 and uB=dB +[-dB-dA]/2 As a result the two agents split the difference: the agents first
agree to take a part of the cake equal to their di and then they split the remaining cake equally between themselves
7Chapter 9
Agent Technology for e-Commerce
Strategic bargaining
In strategic models of bargaining, the bargaining solution emerges as the equilibrium of a sequential game in which the parties take turns in making offers and counteroffers
Two agents A and B bargain about the partition of a cake Offers are made at discrete points in time An offer is a number 0 and At each moment in time each agent makes an offer to the other; if
the other accepts, the game ends, otherwise the game continues with the other agent now making an offer
8Chapter 9
Agent Technology for e-Commerce
The bargaining process is not frictionless: agents are impatient and they would rather agree on the same deal today rather than tomorrow. This is expressed as a discount factor =exp(-ri)
If the agents reach a deal at time point t then agent i’s payoff is oiexp(-rit)
The bargaining situation can be depicted as a sequential game with subgames in extensive form
9Chapter 9
Agent Technology for e-Commerce
A
offer oA
B
accept
[oA , (1-oA)]
reject B
offer oB
A
accept
[A(1- oB), BoB]
rejectA
offer oA
B
accept
[AA oA, B B(1-oA)]
rejectB
Subgame 1
Subgame 2
Subgame 3
…
10Chapter 9
Agent Technology for e-Commerce
The basic alternating offers game has a subgame perfect Nash equilibrium:
Agent A gets (1-B)/(1- A B)
Agent B gets 1 minus (1-B)/(1- A B)
The unique subgame perfect Nash equilibrium satisfies two properties
No delay: whenever an agent has to make an offer, the equilibrium offer is accepted by the other agent
Stationarity: in equilibrium, a player makes the same offer whenever it has to make an offer
11Chapter 9
Agent Technology for e-Commerce
The following strategies define the unique subgame perfect equilibrium
Player A always offers
and always accepts an offer
Player B always offers
and always accepts an offer
12Chapter 9
Agent Technology for e-Commerce
The Strategic Negotiation Protocol
Based on Rubinstein’s protocol of alternating offers N agents A={a1,….,an} need to agree on a given issue They can take actions at certain times T={0,1,..} In each period tT of the negotiation if an agreement hasn’t
been reached, the agent whose turn is to make an offer at time t will suggest a possible solution
Each of the other agents responds by accepting (Yes), refusing (No), or opting out of the negotiation (Opt)
13Chapter 9
Agent Technology for e-Commerce
If all the agents choose Yes then the negotiation ends and the solution/offer is implemented
If at least one of the agents opts out, then the negotiation ends and a default solution is implemented
If no agent has opted out, but at least one has refused the offer, the negotiation proceeds to cycle t+1 and the next agent makes a counteroffer
An agent that responds to an offer is not aware of the other agents’ responses in the current negotiation period
14Chapter 9
Agent Technology for e-Commerce
Assumptions:
1. Rationality
2. Agents avoid opting out
3. Agreements are honoured
4. No long-term commitments
5. Common knowledge. Assumptions 1-4 are common knowledge
15Chapter 9
Agent Technology for e-Commerce
Utility functions
An agent has a utility function over all possible outcomes o The time and resources spent on the negotiation process affect
this utility
Types of utility functions: Fixed losses/gains per time unit: ui(o,t)=ui (o,0)+tci
Time constant discount rate: ui (o,t)= it · ui(o,0) where 0<i
t<1. Every agent i has a fixed discount rate i
t
16Chapter 9
Agent Technology for e-Commerce
Models with a financial system with an interest rate r:
Finite-horizon models with fixed losses per time unit:
ui(o,t) = ui (o,0)(1-t/k)-tc for t k
(applicable when it is known in advance that the outcome is valid for k periods)
17Chapter 9
Agent Technology for e-Commerce
The SNP is useful in situations where: Agents do not agree on any entity-oracle who may provide a
centralized solution The system is dynamic and therefore a predefined solution
cannot be imposed A centralized solution may cause a performance bottleneck There is incomplete information and no entity-oracle has all the
relevant information
Applications: data and ask allocation, negotiation over pollution issues, hostage negotiation
Applications of the SNP
18Chapter 9
Agent Technology for e-Commerce
Negotiation in different domains
Two broad categories: Task-oriented domains Worth-oriented domains
19Chapter 9
Agent Technology for e-Commerce
Task-oriented domains (TOD): an agent’s activity can be defined in terms of a set of tasks, where a task is a nondivisible job
Example A has to post letters and return a few books to the library B has to post a package and visit the library to borrow this
month’s National Geographic Both agents could benefit if they could reach an agreement
Negotiation in task-oriented domains
20Chapter 9
Agent Technology for e-Commerce
Task-oriented domains
A task-oriented domain can be formalized as a tuple T,A,c: T is a finite set of tasks A is the set of agents and any agent is capable of carrying out any
combination of tasks c is the cost function which takes as parameters the set of tasks;
c(T’) is independent of which agent carries the tasks in list T’
21Chapter 9
Agent Technology for e-Commerce
An encounter within a TOD is an ordered list of tasks T1,…,Tn such that Ti is the list of tasks allocated to agent ai
A deal = D1,D2 is an allocation of tasks T1T2
The cost of a deal to agent ai will be denoted costi() and the agent’s utility is:
ui()=c(Ti)- costi() If the agents fail to agree on a deal, a default conflict deal is
implemented and ui()=0 A Pareto efficient allocation or deal cannot be improved upon by
any of the agents without making any other agent worse off
22Chapter 9
Agent Technology for e-Commerce
Monotonic concession protocol
The negotiation proceeds in rounds: In round 1, both agents propose a deal from the negotiation set
simultaneously An agreement is reached and the protocol terminates when one of
the agents finds that the deal proposed by the other is at least as good or better than its own proposal
If no agreement is reached, the negotiation proceeds to the next round
23Chapter 9
Agent Technology for e-Commerce
In round t+1, both agents make proposals: A new proposal can be the previously made proposal by the agent
(the agent stands still), or A new proposal which gives the other agent more utility than the
proposal made in round t (the agent concedes) If none of the agents make a concession, the protocol terminates
with the conflict deal
24Chapter 9
Agent Technology for e-Commerce
A’s best deal B’s best deal Conflict deal
Maximal loss from concession
Maximal loss from conflict deal
25Chapter 9
Agent Technology for e-Commerce
The Zeuthian strategy
Three aspects: What should an agent’s first proposal be?
The best deal for that agent
Who should concede on any given round?
The agent that has more to loose if the conflict deal is imposed
If an agent concedes, how much should it concede?
As much as it is required so that the balance of risk is changed between the agent and its opponent
26Chapter 9
Agent Technology for e-Commerce
Measuring the degree of willingness to risk Suppose A has conceded a lot already, then the deal is very close
to the conflict deal and A does not have much to loose The extent to which an agent is more willing to risk conflict is:
As dwriski,t increases, the agent has less to lose if a conflict occurs and as a result will not be willing to concede
The agent with the lowest dwriski,t should concede
27Chapter 9
Agent Technology for e-Commerce
The agents will not run into conflict, i.e. the outcome reached is going to be Pareto efficient
Not in Nash equilibrium, a self-interested agent knowing that the opponent is using the Zeuthian strategy can try and exploit this
Extended Zeuthian strategy: who concedes in case both agents have the same dwriski,t is decided on the flip of a fair coin
This is now a game where the players play with mixed strategies, so there is at least on mixed strategies Nash equilibrium
But there is some positive probability that the conflict deal will be reached. So although the extended Zeuthian strategy is stable, it may yield an inefficient outcome
Not computational and communication efficient
Features of the Zeuthian strategy
28Chapter 9
Agent Technology for e-Commerce
Deception in TODs
Agents have to declare their tasks, and may do so insincerely An agent can declare phantom or decoy tasks in an attempt to
influence the outcome of the negotiation process. If an agent can produce a phantom task on demand then this is
called a decoy Phantom tasks that cannot be easily produced make deception
detection easier An agent can also hide tasks
29Chapter 9
Agent Technology for e-Commerce
Worth Oriented Domains
Agents are interested in bringing about states that have the greatest value
Agents’ goals can be achieved through joint plans
30Chapter 9
Agent Technology for e-Commerce
Worth-oriented domains can be formalized as a tuple S, A, J, c: S is the set of all possible states A is the set of agents J is the set of all possible joint plans c is the cost function which represents the cost of a joint plan to
an agent ai
j:s1|→s2 denotes that the execution of plan j is s1 leads to s2
If the agent were alone in the world, then its utility from bringing the world to its own ‘stand-alone optimal’ using its own plan is:
31Chapter 9
Agent Technology for e-Commerce
It may be impossible for each of the agents to perform single-agent plans to bring the world to a desirable state
Agents in WODs can reach a compromise by negotiating not only over what parts of their goals will be achieved, but also over the means
State-oriented domains: the worth value is associated only with the achievement of an agent’s full goal
32Chapter 9
Agent Technology for e-Commerce
Coalitions
A coalition is a set of agents that agree to cooperate in order to achieve a common objective
The incentives for creating/joining a coalition can be: Monetary: reduction of cost or increased profit Risk reduction (or allowing someone else to assume risk) Increase in market size or share
33Chapter 9
Agent Technology for e-Commerce
Coalition formation
Coalition formation can be studied in the context of characteristic function games (CFG):
A set N of agents in which each subset is called a coalition The value of a coalition S is given by a characteristic function vS
CS: the coalition structure is the set of all coalitions such that every agent belongs to one
The solution of a game with side payments is a coalition configuration which consists of a partition S of N, the coalition structure CS, and an n-dimensional payoff vector
34Chapter 9
Agent Technology for e-Commerce
Coalition formation in CFG games involves two activities: Coalition structure generation Division of the value of the generated coalition structure among
all agents
The two activities are intertwined
35Chapter 9
Agent Technology for e-Commerce
Coalition structure generation
The formation of an optimal, maximum welfare coalition structure is trivial when the coalition values are:
Super-additive: there is at least one optimal coalition structure, the grand coalition
Sub-additive: the optimal coalition structure is the one in which every agent acts on their own
When games are neither sub-additive or super-additive some coalitions are best off merging whereas others are not
36Chapter 9
Agent Technology for e-Commerce
The objective is to maximize the social welfare of the agents by finding an optimal coalition structure CS*:
where V(CS) is the value of a coalition structure:
37Chapter 9
Agent Technology for e-Commerce
The number of coalition structures CS is exponential in the number of coalitions S, the agent must search among O(nn) coalition structures to find the optimal one
The number of coalitions is
Not all coalition structures can be enumerated unless n is small Can the agents approximate the optimal coalition structure? Can they search through a subset LM such that:
38Chapter 9
Agent Technology for e-Commerce
Coalition structures for four agents
• The lowest two levels of the ordering (j=1 and j=2) the agents have seen all the possible coalition structures• The agents must at least inspect 2n-1 different coalition structures in order to determine a worse-case bound• If more time for computation is available more coalition structures can be inspected
39Chapter 9
Agent Technology for e-Commerce
Division of payoffs
Payoff division is important as it affects the stability of the coalitionMany coalition formation algorithms rely on game theory concepts
The Core
The strongest solution concept; it may be empty Agents may switch indefinitely between coalitions The Core may contain multiple solutions – the agents need to
agree on one: the nucleolus Calculating the Core is an NP-hard problem
40Chapter 9
Agent Technology for e-Commerce
The Shapley value: Agent i is a dummy if vSi-vS=vi for every coalition S that does not
include i Agents i and j are interchangeable if for all S with either i or j, vS
remains the same if i is replaced by j
We require a set of payoffs that satisfy: Symmetry: if i and j are interchangeable then pi=pj
Dummies: if i is a dummy, then pi=v{i}
Additivity: for two games v and w, pi in v+w is equal to pi in v plus pi in w
41Chapter 9
Agent Technology for e-Commerce
The Shapley value satisfies these conditions and sets the payoffs to
It always exists and is unique Pareto efficient It guarantees that individual agents and the grand coalition have
an incentive to stay with the coalition structure No guarantee that all subgroups of agents are better off in the
coalition structure than by splitting out into a coalition of their own
42Chapter 9
Agent Technology for e-Commerce
Customer coalitions
Suppose you want to buy a PC, you can do so at retail price If nine of your friends are interested in the same type of PC, you
can join forces and ask retailers to make you a better offer as this is a bulk purchase
What the discount is depends on the number of PCs The vendor has an incentive to lower the price, as otherwise the
sale will be lost
43Chapter 9
Agent Technology for e-Commerce
Supplier incentive to sell wholesale
Utility to sell wholesale:
The utility of selling n items retail:
The utility of selling n items wholesale:
Up to some number nretail, the supplier does not have an incentive to sell wholesale as marketing costs are identical
44Chapter 9
Agent Technology for e-Commerce
Customer incentive to buy wholesale
A customer’s utility:
ucustomer = vitem – pitem – cstorage
Maximum utility range: MUR(nmin,nmax) – utility is high while the management or storage costs remain low
If nwholesaleMUR then the customer can purchase the items at wholesale price
But the customer needs to be given incentives to buy larger quantities, i.e, the supplier needs to lower the price
45Chapter 9
Agent Technology for e-Commerce
In practice, individual consumers very rarely require large enough quantities so that they can purchase at wholesale prices
But by forming coalitions, consumers can increase the quantity purchased so as to be charged wholesale prices
The utility of the coalition is now MURcoalition = MURi
If nwholesaleMURcoalition then the coalition can make a wholesale purchase
46Chapter 9
Agent Technology for e-Commerce
Coalition protocols
The general stages involved in a coalition protocol are: Negotiation: The coalition leader/representative negotiates with
suppliers Coalition formation: The initiator/leader invites potential
members to join the coalition; possible admission constraints Leader election/voting: The members may elect a leader. Not all
protocols have this stage Payment collection: The coalition leader/representative collects
payments and pays supplier. Execution/distribution: The transaction is executed; the goods
arrive and they are distributed to the members of the coalition
47Chapter 9
Agent Technology for e-Commerce
Issues in coalition protocols
Coalition stability Distribution of utility and costs Trust
Negotiation stage Payment collection stage Distribution stage
Distribution of risk Risk of transaction failure Risk of coalition failure Price uncertainty
48Chapter 9
Agent Technology for e-Commerce
Coalition protocols
Assume a coalition leader (L), a set of suppliers S={s1,s2,…,sk} and a set of potential coalition members M={m1,m2,…,mn}
Based on the order in which the negotiation and coalition formation stages take place there are two types:
Post-Negotiation Pre-Negotiation
49Chapter 9
Agent Technology for e-Commerce
Post-negotiation protocol
1. LCS: L advertises the creation of a coalition with certain parameters (deadline, maximum number etc.)
2. Each miM considers whether to join the coalition and sends necessary message mi L: “Join the Coalition”
3. At the expiration of the coalition deadline/size limit, the leader enters the negotiation with the suppliers si S using its private protocol/strategy and decides on a deal
4. L collects money from group members, and arranges for the shipping and distribution of goods
50Chapter 9
Agent Technology for e-Commerce
Issues Trust in the coalition leader is required Shills can start coalitions
Trust can be established in a number of ways Leaders can be elected A trusted third party can be appointed to conduct the negotiations The coalition leader could be compelled to open every step of the
negotiation to the scrutiny of group members Members can vote on the suppliers’ bids – but time-consuming
51Chapter 9
Agent Technology for e-Commerce
Pre-negotiation protocol
1. L conducts negotiations with the suppliers S, using its private parameters.
2. L opens the coalition to potential coalition members, disclosing the details of the deal agreed
3. Each miM considers whether to join the coalition and sends necessary message mi L: “Join the Coalition”
4. After a certain period of time elapses, or the coalition gains some minimum number of members, L closes the coalition to new members and executes the transaction
52Chapter 9
Agent Technology for e-Commerce
Issues An insufficient number of members join the coalition The deal must be re-negotiated, resulting in a higher price and,
possibly, more members leaving the coalition
53Chapter 9
Agent Technology for e-Commerce
Variation: L presents not an estimated group size, but a range of sizes. The supplier bids with a step function P = Fbid(quantity) The risk in the transaction is shifted onto the coalition members
due to the price uncertainty A buyer’s decision on whether to join depends on its estimate of
the probability that the final price will be lower than its preservation:
pmax-coalition >= preservation >= pmin-coalition
A dominant strategy for a buyer would be to wait until the coalition is almost closed for new members
54Chapter 9
Agent Technology for e-Commerce
Distribution of costs and utility
The coalition leader can operate on: Non-Profit: ccoalition is distributed either equally among all
participants or on the sub-lot size basis. Can form on a per need basis or be stable ‘buyer's clubs’ that exist over time
For-Profit: Consolidator: Pre-negotiates a deal with the supplier given an
estimated group size, and then re-sells the items individually, keeping enough of the savings to cover ccoalition and profit
Rebater: Sells the items at retail price minus a small rebate, and keeps the rest of the savings
55Chapter 9
Agent Technology for e-Commerce
Social choice problems
Given a society of agents and their preferences we would like to aggregate them into a social ‘preference’
How can we go from often divergent and incompatible but individually consistent views on what is the socially best outcome, to a single and socially consistent view?
56Chapter 9
Agent Technology for e-Commerce
Social choice rule
A social choice setting: N: a set of agents (society) : a set of feasible outcomes for the society iN there is an asymmetric and transitive preference relation
on
Social choice rule takes as input the agents’ preference relations
and produces as output the social preferences denoted by a relation
57Chapter 9
Agent Technology for e-Commerce
Conditions: A social preference ordering should exist for all possible
inputs and should be complete and transitive over The outcome should be Pareto efficient The scheme should be independent of irrelevant alternatives No agent should be a dictator: no o o’ implies o o’ for all
preferences of the other agents
Arrow’s Impossibility Theorem:
There is no social choice rule that satisfies all four conditions
58Chapter 9
Agent Technology for e-Commerce
Voting protocols
A class of social choice rules: the third condition is relaxed Agents give input to a mechanism and the mechanism chooses
an outcome based on these inputs which is the solution imposed upon all participating agents
The aim is to enhance the general good (social welfare) Binary protocols Plurality protocols Mixed protocols
59Chapter 9
Agent Technology for e-Commerce
Binary protocols
Agents are asked to choose between two alternatives at a time, if there are more than two, these are compared pairwise and the winner challenges further alternatives
Condorcet protocol: each alternative is pitted against all other others and the one that defeats all others wins
They may not generate a transitive social preferences ordering:
60Chapter 9
Agent Technology for e-Commerce
The outcome depends on the agenda:
28% prefer c d b a
25% prefer a c d b
24% prefer b a c d
23% prefer a d c b
a, b, c, d b, d, c, a c, a, d, b
c
b
c
b
d
a
c
c
d
c
d
a
b
a
a
a
b
d
a
c
b
c, a, b, d
b
a
b
a
d
c
d
61Chapter 9
Agent Technology for e-Commerce
Plurality protocols
All alternatives are compared simultaneously The winner is the alternative with the highest number of votes Such protocols are used in political elections
62Chapter 9
Agent Technology for e-Commerce
Borda protocol
The Borda count assigns an alternative | | points whenever it is highest in some agent’s preference list | |-1 whenever it is second and so on
The alternative with the highest count becomes the social choice But, it can lead to paradoxical results if an irrelevant alternative
is removed
64Chapter 9
Agent Technology for e-Commerce
Mixed protocols
Some protocols combine plurality and binary protocols
Majority runoff protocol First stage: voters indicate their preferences among a set of
alternatives by casting one vote. If an alternative receives the majority of votes, this is the winner. Otherwise:
Second stage: the two most preferred alternatives run against each other
Proportional representation The full preference rankings of the voters provide for a
proportional representation
65Chapter 9
Agent Technology for e-Commerce
Variation of proportional representation: single transferable vote or Hare protocol:
Agents cast one vote, but indicate their preference rankings over all alternatives
Alternatives which obtain a certain percentage of votes are elected and those that fail to obtain that percentage (or the alternatives with the fewer votes) are eliminated
The votes from the eliminated alternatives are transferred to the next highest ranking alternative according to the agents’ preference rankings
The processes is repeated until an appropriate number of alternatives is elected
66Chapter 9
Agent Technology for e-Commerce
Issues in voting protocols
Agents may declare their preferences insincerely or vote strategically in order to increase their own gain
Those responsible for setting up the process may attempt to manipulate the proceedings
The application of different protocols to the same situation may lead to different outcomes
67Chapter 9
Agent Technology for e-Commerce
A function that aggregates the individual agents’ preferences into a social one is called a social welfare function
Summing up the agents’ utilities according to an allocation is such a social welfare function
The allocation o is preferred to o’ if:
Weighted sum of utilities
Social welfare functions
68Chapter 9
Agent Technology for e-Commerce
Welfare maximization
The social welfare maximization problem takes the form:
such that
Such a maximal welfare allocation is Pareto efficient
69Chapter 9
Agent Technology for e-Commerce
Argumentation
Negotiation as purported by game theory has two limitations: Proposal/offers and negotiation positions cannot be justified Proposals/offers and negotiation positions cannot be modified
These limitations can be overcome through argumentation-based negotiation
Additional information can be exchanged on top of the offers Agents enter into a dialogue and attempt to convince the others
(persuasion type of dialogue according to Walton and Krabbe)
70Chapter 9
Agent Technology for e-Commerce
Argumentation amongst humans: Logical mode: resembles logical or mathematical proof Emotional mode: makes use of one’s feelings, emotions and
other attitudes Visceral mode: such arguments involve physical and social
aspects Kisceral mode: makes appeal to the religious, mystical or
intuitive side of human nature
71Chapter 9
Agent Technology for e-Commerce
Generating persuasive arguments
Two parties: the persuader and the persuadee Persuasive arguments are used in order to change the behaviour
of the pursuadee – behaviour changes not necessarily the beliefs
Arguments may: explain the opinion of the agent on a particular proposal or
provide a critique on a proposal which explains why it is unacceptable – the negotiation space of the agent is explained
give reasons why the agent should accept a proposal – attempt to convince the other party about the validity of a proposal
72Chapter 9
Agent Technology for e-Commerce
To generate persuasive arguments, agents need to be able to: Represent and maintain the beliefs of other agents Select which beliefs need to be influenced and in what way Connect beliefs with behaviour Choose the most appropriate and convincing argument Offer counter-arguments Modify one’s own position as the dialogue process progresses
A complex cognitive task
73Chapter 9
Agent Technology for e-Commerce
Abstract architecture
Message interpretation
Message generation
Communication component
Knowledge base component
Environment model
Self-model
Opponent model
Negotiation history
Decision making component
ProposalProposalhistory
UpdateQuery
Proposal
Message
Message
Argument evaluation
Argument
Argument generator
Argument selector
Update
Argumentation component
Arguments
Response
74Chapter 9
Agent Technology for e-Commerce
The PERSUADER system
Domain: negotiations between union and employers The system plays the role of the mediator Three main tasks
Proposal generation Counter-proposal generation based on proposal from the
disagreeing participant Persuasive argumentation
Agents have a representation of the others’ beliefs which they update based on the proposals and arguments made during the negotiation process
75Chapter 9
Agent Technology for e-Commerce
The system can generate different types of arguments (in order of increasing convincing power):
Appeal to universal principle Appeal to a theme or a package of goals Appeal to authority Appeal to status quo Appeal to ‘minor standard’ Appeal to prevailing practice Appeal to precedents and counterexamples Appeal to self-interest Threats and promises
76Chapter 9
Agent Technology for e-Commerce
Logic-based argumentation
Generating a series of logical arguments for and against propositions, offers etc.
A logical argument is of the form :
where is a knowledge base containing facts about the world (which
may not necessarily be consistent is the sentence (offer, position etc.) that is to be proved, i.e. the
conclusion KB is a set of logical formulas that
And can be proven from KB
77Chapter 9
Agent Technology for e-Commerce
Given an argument (, KB) there are two types of argument against it:
Those that rebut it: argument (1, KB1) rebuts (2, KB2) if 1 attacks 2
Those that undercut it: argument (1, KB1) rebuts (2, KB2) if 1 attacks for some KB2
Attack: for any propositions and , attacks , if an only if
78Chapter 9
Agent Technology for e-Commerce
Classes of arguments (in order of increasing acceptability): A1: all arguments that can be made from A2: all nontrivial arguments that can be made from A3: all arguments that can be made from for propositions
for which there are no rebutting arguments A4: All arguments that may be made from for propositions
for which there are no undercutting arguments A5: All tautological arguments
79Chapter 9
Agent Technology for e-Commerce
Negotiation as dialogue games
Dialogue games describe interactions among agents where each agent ‘makes a move’ by making an utterance according to a set of rules
Commencement rules Locution rules Combination rules Commitment rules Termination rules
80Chapter 9
Agent Technology for e-Commerce
A dialogue game between buyers and sellers
The dialogue game of McBurney et al. (2003) includes 7 stages:
1. Open dialogue
2. Inform
3. Consideration set formation
4. Option selection
5. Negotiation
6. Confirmation
7. Dialogue termination
Strictly speaking stages (3) and (4) are not part of the dialogue
81Chapter 9
Agent Technology for e-Commerce
The simplest form of dialogue consists of all 7 stages More complex dialogues can be formed by entering some stages
multiple times subject to the following rules: The first stage must be the Open dialogue and occurs once The last stage is the Dialogue termination and occurs once Every dialogue that terminates normally involves the Open
dialogue and Dialogue termination stages The first instance of every other stage apart from the first and
last one must be preceded by an instance of the Inform stage The Confirmation stage may only be entered following an
instance of the Negotiation stage
82Chapter 9
Agent Technology for e-Commerce
To automate dialogues appropriate locutions are required: Open dialogue: open_dialogue(.) followed by at least one
enter_dialogue(.)
Inform: seek_info(.) and willing_to_sell(.)
Negotiation: desire_to_buy(.), prefer(.), refuse_to_buy(.) and refuse_to_sell(.)
Confirmation: agree_to_sell(.) and agree_to_buy(.)
Dialogue termination: withdraw_dialogue(.)