740 likes | 855 Views
Agent-mediated Interaction. From Auctions to Negotiation and Argumentation Carles Sierra IIIA-CSIC Barcelona Utrecht, October 13, 2000. IIIA-CSIC. Talk plan. Auctions: FISHMARKET Negotiation Argumentation Robot navigation Electronic Institutions. Introduction.
E N D
Agent-mediated Interaction. From Auctions to Negotiation and Argumentation Carles Sierra IIIA-CSIC Barcelona Utrecht, October 13, 2000 IIIA-CSIC
Talk plan Auctions: FISHMARKET Negotiation Argumentation Robot navigation Electronic Institutions
Introduction Agents inhabiting the same environment need to co-ordinate their activities to improve their individual or collective performance. The aim of DAI is to design intelligent sistems that behave efficiently. A common assumption in many applications, specially in AMEC, is that agents are self-interested and utility maximisers. In others, agents are co-operative. DAI is divided in two big areas: Distributed problem solving, where the designer determines the protocol and the strategy (relation between state and action) of each agent, and Multi Agent Systems, where the agents are provided with an interaction protocol but chose the strategy to follow.
Auctions Auctions are mechanisms very frequent in MAS. They have been deeply analysed by economists. There are three types: 1) Of private value, e.g. a cake. 2) Of common value, e.g. treasure bonds. 3) Of correlated value, e.g. contracts. Protocols: English. If it is of private value, the strategy is to increase the bids until the reserve price. In those of correlated value the auctioneer may increase the price in predetermined amounts. Sealed bid. There is no dominant strategy. Dutch. Equivalent to sealed bid. They are very efficient. Vickrey. The dominant strategy is to bid for the reserve price.
Auctions: the Fishmarket Seller’s admitter
Implementation LAN Auditor LLotja virtual Modems Cap Admissió de venedors Servidor Admissió de compadors Admissió de peix Gestió de compadors Gestió de venedors Subhastador Interagent comprador Interagent venedor Agent venedor Agent comprador
FM 1.0: A test-bed for Electronic Auctions • Realistic.Grown out of a complex real world application. • Multi-user • Architecturally neutral • Customizability and repeatibility • Agent-builder facility (Library of agent templates) • Monitoring and Analysis facilities • Market scenarios as tournament scenarios.
Bargaining In bargaining, agents may make deals that are mutually beneficial, but they are in conflict over which deal to chose. Negotiation mechanisms fall mainly on strategic bargaining. Axiomatic Theory. The desired solutions are not those found in a certain equilibrium, but those that satisfy a set of axioms. Classical axioms are those of Nash: outcome u*=(u1(o*), u2(o*)) must satisfy: Invariance: The numerical utilities of agents represent ordinal preferences, numerical values don’t matter. Thus, the utility functions must satisfy that for any f linear and increasing: u*(f(o), f(ofail))=f(u*(o, ofail)) Anonimity: Changing the labels of the players does not affect the outcome. Independence of irrelevat alternatives: if we eliminate some o, but not o*, o* is still the solution. Pareto eficiency: we cannot give more utility to both players over u*=(u1(o*), u2(o*)).
Bargaining Strategic Theory: No axioms on the solution are given, the interaction is modelled as a game. The analysis consists on finding which strategies of the players are in equilibrium. It explains the behaviour of utility maximisers better than the axiomatic theory (where the notion of strategy does not make much sense). The theory of negotiation is basically here. Without assuming perfect rationality, the computational costs of the deliberation and the potential benefits of bargaining conflict. AI has many things to say on this task.
Negotiation • Commerce is about interaction • Between buyers and sellers at all stages: finding, purchasing, delivery. • First generation • Passive web query • Simple interactions: auctions • Second generation • Rich and flexible interactions • Negotiation is the key type of interaction • Process by which groups of agents communicate with one another to try and come to a mutually acceptable agreement on same matter. • Many forms exist: auctions, contract net, argumentation. • It is key because agents are autonomous: an acquaintance needs to be convinced to be influenced. • Negotiation is achieved by making proposals, trading options, offering concessions.
Negotiation components • Negotiation objects. Issues of the agreements. Number of them, types of operations on them. • Negotiation protocols. Rules that govern the interaction: permissible participants, valid actions, negotiation states. • Agents reasoning model. Decision making apparatus. From simple bidding to complex argumentation. • Challenges • Trust • Protocol engineering • Reasoning models
Negotiation object example Real State Agency. Seller b and buyer a. Issues={Address,Surface,Rooms,Brightness,Price,Garage} Negotiation thread:
Negotiation reasoning model Each agent a negotiates over a number of issues that have a: 1) Delimited range [minj, maxj] 2) Monotonic scoring function Vja: [minj, maxj]-> [0,1] 3) Relative importance, wja The utility function for an agent a has the following form: The negotiation protocol consists of an iterative process of offers and counteroffers until a deal is reached.
Price:2 Quality:5 A B ? Price:9.9 Quality:1.1 Tactic: trade-offs
Trade-off Mechanism (I) • Trade-off is lowering of utility on some issues and simultaneously demanding more on others. • Steps: given x (a’s offer) and y (b’s offer) • (1) Generate all / subset of contracts with the same utility () • isoa() = {x | Va(x) = } • (2) selection of a contract (x´) that agent abelieves is most preferable by b. • Ba (Ub(x´) > Ub(x)) • Ua(x´) + Ub(x´) > Ua(x) + Ub(x) (maximization of joint utility) • Ua(x) = Ub(x´) • Step (2) is an uncertain evaluation: must model Ba
Fuzzy Similarity • Select a contract from isoa() = {x | Va(x) = } that is “closest” or most similar to y. • Implications of this choice: • not the probable choice of the other, but rather, the closeness of two contracts • Not modeling of others but the domain • need a logic of degrees of truth (Zadeh) as opposed to binary truth values of true or false
Definition of Similarity • Sim( ) defined as: Sim(x,y) = j J wj Simj(xj,yj) Simj(xj,yj) = 1i m(hi(xj) hi(yj)) • where wjis the agent´s belief about the importance the other places on each issue in negotiation • hi( ) is ith comparison criteria function (e.g warmth) • is the conjunction operator (e.g minimum) • is the equivalence operator (e.g 1-| hi(xj)-hi(yj)|)
An Example of Similarity • Dcolours{yellow,orange,green,cyan,red,...} • Similarity of colours according to different perceptive criteria: • Temperature (warm v.s cold colours) • Luminosity • Visibility • Memory • dynamicity ht = {(yellow, 0.9), (violet, 0.1), (magenta, 0.1), (green, 0.3), (cyan, 0.2), (red, 0.7),...} hl = {(yellow, 0.9), (violet, 0.3), (magenta, 0.6), (green, 0.6), (cyan, 0.4), (red, 0.8),...} hv = {(yellow, 1), (violet, 0.5), (magenta, 0.4), (green, 0.1), (cyan, 1), (red, 0.2),...}
Similarity of Colours • Simcolour(yellow, green) = min( 1- |ht(yellow)- ht(green)|, 1-| hl(yellow)- hl(green)|, 1- |hv(yellow)- hv(green)|)= min(0.4,0.7,0.1) = 0.1 • Simcolour(yellow, red) = min( 1- |ht(yellow)- ht(red)|, 1-| hl(yellow)- hl(red)|, 1- |hv(yellow)- hv(red)|)= min(0.8,0.9,0.2) = 0.2 • yellow is more similar to red than to green on these criteria • sim(yellow,green) and sim(yellow,red) • simcolour(colour,colour) = 1i m(hi(xcolour) hi(ycolour)) • i={temperature,luminosity,visibility}
y y ? x X´ x The Trade-off Algorithm To be beneficial to the other the preference of the other must match the similarity function trade-offa(x,y) = arg maxz isoa() {Sim(z,y)} complexity kn
Argumentation • Autonomy leads to negotiation and to argumentation. • Many problems cannot be solved by a simple offer/counter offer negotiation protocol. • When arguing, agent offers may include knowledge, information, explanations. • The dialogue includes critiques on each others proposals. • Agents must be able to generate arguments as well as rebutting and undercutting other agents’ arguments. • Which argument to prefer may depend on logical criteria or on social considerations. • A logically-based approach to building agents seems natural.
-> + + Hang Mirror -> Hang Picture + + Hang Picture Hang Mirror Hang Mirror -> + + S S B A
-> + + Hang Mirror -> Hang Picture + + Hang Mirror -> + + S S I know agent B has a nail B A
-> + + Hang Mirror -> Hang Picture + + Hang Mirror -> + + S S B A ?
-> + + Hang Mirror -> Hang Picture + + Hang Mirror -> + + Hang Mirror -> + + S S B A
-> + + Hang Mirror -> Hang Picture + + Hang Mirror -> + + Hang Mirror -> -> + + + + Hang Mirror S S B A S S