Intelligent agents

Conflict resolution

in multi agent systems conflicts arise when agents are self interested and the agents represent different stake holders

conflict resolution is possible when there is mutual benefit to reach an agreement

approaches- auctions, voting and negotiations

auctions- to allocate scarce resources

clearly defined protocols, involves third party like auctioneer, uses continuous resource like money, exploits competition between agents

voting- used for group based decisions

single decision from finite options, each agent can have a different preference given by a preference order, defined protocol

negotiation-

less structured protocol, more flexible compared to others, typically involve exchanging offers, often bilateral between two agents, more complex agreements, often decentralised can involve mediator sometimes when multi parties are involved

a deal should be better than having no deal at all, pareto efficiency- there should be no money/resource left, fair agreement

single issue bargaining-distributive

multi issue-integrative

protocols-

ultimatum game- take it or leave it method, agent 1 offers agent 2 can accept or reject it, first person advantage

alternating offers- agent exchanges offers in an alternating manner, 1 offers 2 can either accept reject and counter offer or break off the negotiation, there is a deadline and the last offer resembles the ultimatum game, last offer advantage

monotonic concession- negotiation proceeds in rounds, simultaneously propose offers without seeing the other persons offer, both match one gets chosen or else next round if neither concede by last round it ends without a deal

divide and choose- agent chooses a portion/percentage etc and agent 2 chooses one of those portions, works for resource allocation option

agent preference over outcomes, utility function U(o) where o£O O=offers

example: in cardinal utility functions-

U[seller](p)=p-r

U[seller](disagreement)=0

U[buyer](p)=v-p

U[buyer](disagreement)=0

p=price, r=reserve value (minimum price willing to sell), v=value(max willingness to pay)

multi issue negotiation

offers multiple factors like time, quality, service, delivery time etc

additive utility function= sum of utilities*weights

utility space

U1(o)=w1*o1+w2*o2

U2(o)=w2*9(1-o1)+w2*(1-o2)

pareto efficient agreement

pareto efficiency- no further improvement is possible in utility of one agent without reducing the utility of other agent

fairness- utilitarian (max the sum of utilites), egalitarian(max the min utility), nash bargaining solution (max the product of utility of agents-disagreement payoff), envy freeness(no agent prefers the other agents resources to theirs)

envy freeness-

agent is envious if U1(1-o1,1-o2)>U1(o1,o2)