Automated Negotiations Yelena Yesha Olga Streltchenko 1 Automated

Automated Negotiations Yelena Yesha Olga Streltchenko 1

Automated Negotiations z. Issues in automated negotiations z. Auctions overview z. Electronic auctions (e-commerce servers) 2

Negotiation in electronic commerce z Negotiation: key component in e-commerce z Two or more parties multilaterally bargain resources for intended gain, using the tools of electronic commerce, ye. g. agents negotiating a solution electronically. z Negotiating function is performed through (networked) computers. 3

Automated Negotiation z. Auto: negotiation is performed by computational agents, which y. Represent real-world parties; y. Perform information retrieval and processing; y. Find & prepare contracts; y. Perform other activities. 4

Electronic marketplace z Designated “meeting place” for negotiating parties. z A trusted intermediary that facilitates trading between buyers and sellers on the Web. y``Closed'' marketplace: xpredefined set of users; xenrollment. y``Open'' marketplace: xagents enter and exit any time. 5

Levels of automation z. Negotiation support systems. y. Help human negotiations. z. Intelligent Agents. y. Negotiate electronically within an environment governed by rules. y. No human intervention. 6

Automation z. Architecture Issues for Automated Markets z. Transaction processing. z. Decision support. 7

Architecture Issues for Automated Markets: Banking System z E-Market entities (negotiating agents, etc. ) have to be able to communicate with the existing banking and financial services. y. Integration of marketplace interfaces into banking legacy systems. z Open standard for agent-to-bank communications. z Security. 8

Architecture Issues for Automated Markets: Communication Infrastructure z Efficiency and robustness. z Redundancy y. E. g. a mesh of redundant hubs interconnected with each other. z Open standard communication systems to allow development of platform-independent systems that plug into a marketplace architecture. z Independent of agent architectures. z Agents must yaccess global posting services; yuse common language for outbound communications. 9

Standardized Communication Infrastructure z. Integration into the back ends of electronic catalog databases and existing EDI systems. z. Common language for outbound communications; y. KQML. 10

Architecture Issues for Automated Markets: Transfer and storage of goods z. Representation and handling of physical goods. y. Goods as software objects. y. Copy-protection to insure that an object is at one place at a time: xencoding by an owner; xaccess by agents authorized by the owner. y. Arrangement for physical shipment. 11

Architecture Issues for Automated Markets: Handling of Electronic Items z. Format of a software object. z. Copy-protection as above. z. Delivery channels. 12

Architecture Issues for Automated Markets: Administration and Policies z. Central administration to provide ydefault protection; yprevention of illegal transactions; ycollection of taxes and commissions; ycredit and service ratings for agents. 13

Self-interested Agents (SI) vs Cooperative (Distributed) Problem Solving (CDPM) z Self-interested agents act according to their internal considerations and pursue their private goals; y. E. g. internal utility function maximization. y. Competitive. z A multi-agent system may strive to achieve a global (societal) goal; y. E. g. global utility function maximization. y. Requires cooperation on the part of individual agents. y. Common goal in a distributed system. 14

Transaction Processing: Levels of commitment z. For SI agents, contracts are bounding. z. In CDPS commitments are allowed to be broken unilaterally based on some local reasoning. y. Continuous levels of commitment based on a monetary penalty method. 15

Transaction Processing: Decommitting z. Replies vs timeout. y. Inform the other party that the negotiation is not considered any more. 16

Transaction Processing: Message congestion z. Most distributed implementations run into this problem. yhigh risk of saturation. 17

Message Congestion (cont’d) z Remedies: y. Focused addressing x. Heavy load agents with free resources or agents soliciting contracts announce availability; x. Light load agents with tasks or agents offering contracts announce availability. y. Audience restriction x An agent negotiates with a subset of agents in the system. y. Ignoring outdated messages. 18

Decision Support: Learning z. Creating an agent with a complete set of strategies (no learning) vs z. Acquisition of experience from the previous negotiations (learning). x. Genetic algorithms and genetic programming. x. Q-learning (reinforced learning). x. Other techniques. z. Learning is computationally expensive. 19

Decision Support: Perfect Rationality vs Bounded Rationality z Two approaches: y. Microeconomics ( e. g. Kreps~[1990], Varian~[1992] and Raifa~[1982]) y and distributed artificial intelligence (DAI) (Rosenschein and Zlotkin~[1994], Durfee[1994]). z Perfect rationality assumes that an agent can accurately model the environment and perform exact calculation for their decision-making. ycomputation is complex and resource/time consuming. 20

Perfect Rationality vs Bounded Rationality (cont’d) z. Bounded rationality means that resources are costly and bounded, and the model of the environment is not accurate; ye. g. the environment is dynamic and evolves. z. An agent has to ydecide how much computation to perform per task/contract; ychoose a subset of tasks/ contracts available for consideration. 21

Combinatorial Aspects of Negotiation z. Negotiating with several agents or several marketplaces at a time. z. Computation complexity. 22

Automated Negotiations: What's lacking? z. Why does current electronic commerce not widely support negotiations? y. Negotiation is difficult. z. Stumbling blocks. y. Need for a clear unambiguous ontology. y. Need for a strategy. 23

Ontologies z Ontology is a way of categorizing objects such that they are semantically meaningful to a software agent. z Must capture all important attributes of an object to allow for intelligent bargaining on both sides of a negotiation process z Active research area in AI (see, for example, Sowa [1999]). z Existing tools: y. KIF (Knoledge Intergande Format); y. Ontolingua; y. DAML. 24

Strategy z. Game theory y Treats negotiation from a mathematical prospective. y. Recommends a course of actions to a participant taking into account the opponents’ strategies and payoffs. y. Allows formulation of moves beyond pure price determination. 25

Strategy (cont’d) z. Disadvantages of the game-theoretic approach: y. Assumes perfect information identically perceived by all bargaining parties x. Information is asymmetric, expectations are heterogeneous. y. Assumes perfect rationality of all players x. Computational agents operate under constraints, i. e. have bounded rationality. 26

Automated Negotiations: What's lacking? (cont’d) z Sophisticated strategies are mathematically complex and computationally expensive. z Current software agents implement a fairly simple set of governing rules. y. Emerging complexity of the overall system (see, for example, Maes and Chavez [1994]). z Inadequate rule sets might lead to a disastrous system behavior. y. Program trading and 1987 stock market crush. 27

Agent Technology and Automated Negotiations z. Mobility to navigate among electronic marketplaces; z. Intelligence to incorporate sophisticated numerical algorithms for portfolio optimization along with capability to learn and interpret the environment; z. Agency to interact with other market entities (e. g. other agents). 28

Promising application areas z Retail e-commerce z Online auctions z Electricity markets z Bandwidth allocation z Manufacturing planning & scheduling in subcontracting networks z Distributed vehicle routing among independent dispatch centers z Electronic trading of financial instruments 29

Auction - Definitions z An auction is a method of allocating scarce goods, y a method that is based upon competition: x. A seller wishes to obtain as much money as possible, x. A buyer wants to pay as little as necessary. z An auction offers the advantage of simplicity in determining market-based prices. z It is efficient in the sense that it usually ensures that y resources accrue to those who value them most highly; y sellers receive the collective assessment of the value. z The price is set the bidders. 30

When are Auctions Used? z Auctions are useful when y selling a commodity of undetermined quality; y the goods do not have a fixed or determined market value, in other words, when a seller is unsure of the price he can get. z Choosing to sell an item by auctioning it off is y more flexible than setting a fixed price; y less time-consuming and expensive than negotiating a price. z Auctions can be used y for single items such as a work of art; y and for multiple units of a homogeneous item such as gold or Treasury securities. 31

Prices z The price is set not by the seller, but by the bidders. z The seller sets the rules by choosing the type of auction to be used. z The auctioneer doesn't often own the goods, but acts rather, as an agent for someone who does. z The buyers frequently know more than the seller about the value of the item. z A seller, not wanting to suggest a price first out of fear that his ignorance will prove costly, holds an auction to extract information he might not otherwise realize. 32

Bidder Valuations z Reasons for bidding in the auction: y a bidder wishes to acquire goods for personal consumption (wine or art); y a bidder wishes to acquire items for resale or commercial use. z Private valuation y Goods are acquired goods for personal consumption; y The bidder makes his own private valuation of the item for sale. y All bidders have private valuations and tend to keep that information private. x. There would be little point in an auction if the seller knew already how much the highest valuation of an object will be. 33

Bidder Valuations (cont’d) z Common valuation y Goods are acquired goods for resale or commercial use; y An individual bid is predicated not only upon a private valuation reached independently, but also upon an estimate of future valuations of later buyers. Each bidder of this type tries (using the same measurements) to guess the ultimate price of the item. y The item is really worth the same to all, but the exact amount is unknown z Example y Purchasing land for its mineral rights x. Each bidder has different information and a different valuation, but each must guess what price the land might ultimately bring. 34

Taxonomy of Auctions z William Vickrey [Vickrey] established the basic taxonomy of auctions based upon the order in which prices are quoted and the manner in which bids are tendered. He established four major (one sided) auction types: y. English: Ascending-price, open-cry; y. Dutch: descending-price, open-cry, y. First-price, sealed bid, y. Vickrey or second-price, sealed bid. 35

English Auction z The English auction ythe open-outcry auction or the ascending-price auction. z "Here the auctioneer begins with the lowest acceptable price--the reserve price-- and proceeds to solicit successively higher bids from the customers until no one will increase the bid. The item is 'knocked down' (sold) to the highest bidder. ” Paul Milgrom 36

Dutch Auction z In a Dutch auction, bidding starts at an extremely high price and is progressively lowered until a buyer claims an item. z When multiple units are auctioned, normally more takers claim the item as price declines. y. The first winner takes his prize and pays his price y. Later winners pay less. z When the goods are exhausted, the bidding is over. 37

First Price- Sealed Bid z Sealed (not open-outcry like the English or Dutch varieties) and thus hidden from other bidders. z A winning bidder pays exactly the amount he bid. z Usually, (but not always) each participant is allowed one bid which means that bid preparation is especially important. z a sealed-bid format has two distinct periods y a bidding period in which participants submit their bids; y a resolution phase in which the bids are opened and the winner is determined (sometimes the winner is not announced). 38

Multiple Items in a Fist-Price, Sealed Bid Auction z When multiple units are being auctioned, the auction is called "discriminatory" because not all winning bidders pay the same amount. z In a first-price auction (one unit up for sale) each bidder submits one bid in ignorance of all other bids. z The highest bidder wins and pays the amount he bid. z In a "discriminatory” auction, sealed bids are sorted from high to low, and items are awarded at highest bid price until the supply is exhausted. z Winning bidders can (and usually do) pay different prices. 39

The Vickrey Auction z The uniform second-price auction is commonly called the Vickrey auction. z The bids are sealed, and each bidder is ignorant of other bids. z The item is awarded to highest bidder at a price equal to the second-highest bid (or highest unsuccessful bid). y winner pays less than the highest bid. z Example y Suppose bidder A bids $10, bidder B bids $15, and bidder C offers $20, bidder C would win, however he would only pay the price of the second-highest bid, namely $15. 40

The Vickrey Auction (cont’d) z When auctioning multiple units, all winning bidders pay for the items at the same price y the highest losing price. z It seems obvious that a seller would make more money by using a first-price auction, but, in fact, that has been shown to be untrue. z Bidders fully understand the rules and modify their bids as circumstances dictate. 41

Classification English Dutch Seller announces reserve price or some low opening bid. Bidding increases progressively until demand falls. Winning bidder pays highest valuation. Bidder may re-assess evaluation during auction. Seller announces very high opening bid. Bid is lowered progressively until demand rises to match supply. 42

Classification (cont’d) First-price, sealed bid or discriminatory Vickrey Bids submitted in written form with no knowledge of bids of others. Winner pays the exact amount he bid. Bids submitted in written form with no knowledge of the bids of others. Winner pays the second-highest amount bid. 43

Double Auction z In this auction both sellers and buyers submit bids which are then ranked highest to lowest to generate demand supply profiles. z From the profiles, the maximum quantity exchanged can be determined by matching selling offers (starting with lowest price and moving up) with demand bids (starting with highest price and moving down). z This format allows buyers to make offers and sellers to accept those offers at any particular moment. 44

More Auction Varieties z In a sequential auction items are auctioned one at a time. z A continuous double auction is one in which many individual transactions are carried on at a single moment and trading do not stop as each auction is concluded. y Price formation mechanism for financial markets. z In a parallel auction items are open for aution simultaneously and bidders may place their bids during a certain time period. z In a combinatorial auction bidders place bids on combinations of items. z Increase in overall auction complexity is stimulated by rapid development in the areas of y Game theory; y Agent technology; y Electronic commerce (infrastructure). 45

Online Auctions z Spawned by Research Communities y. Auction. Bot (Michigan University); y. Kasbah (MIT Media Lab); ye-Mediator (Tuomas Sandholm et al, Wasington University). z Commercial ye. Bay (www. ebay. com); y. On. Sale (www. onsale. com) - defunct. 46

Auction. Bot (University of Michigan) z General purpose Internet auction server. z Its users create new auctions by choosing from a selection of auction types and its parameters. z User-specified y. Auction type (e. g. English, Dutch, etc. ) y. Parameters x. Clearing times (reserve price, etc. ); x. Method for resolving tie bids; x. Number of sellers permitted. 47

Auction. Bot (cont’d) z Buyers & sellers bid through multilateral distributive negotiation protocols. z Advantage y. Provides an API for users to create own software agents to autonomously compete in Auction. Bot marketplace. z Users encode their own bidding strategies. 48

Kasbah (MIT Media. Lab) z Online multi-agent consumer-to-consumer transaction system. z User that want to buy or sell an item proceed as follows: y. Create an agent; y. Give it some strategic directions; y. Send off to a centralized marketplace. z Proactive y. Agents seek out buyers / sellers and y. Negotiate on behalf of owners; y. Obey users’ constraints. 49

Kasbah (cont’d) z Goal y. Complete acceptable deal on behalf of a user y. Subject to set of user constraints x. Initial bidding (asking) price; x. Lowest (highest) acceptable price; x. Date to complete; x. Restrictions on parties to negotiate with; x. Price change over time. 50

Negotiation in Kasbah z After buying agents and selling agents are matched y. Buying agents offer bid; x. No restriction on time or price. y. Selling agents respond with binding “yes” or “No”. 51

Kasbah Negotiation Strategies z Anxious Seller Buyer y. Linear z Cool-headed y. Quadratic z Frugal y. Exponential z for increasing/decreasing its bid for a product over time. 52

Kasbah: Trust & Reputation (“BBB”) z Better business bureau. z Upon transaction completion both parties may rate other party, e. g. , y. Accuracy of product condition; y. Completion of transaction. z Agents use accumulated ratings y. Determine agents of owners who fall below specified reputation threshold ==> no negotiation 53

e. Mediator: Features z http: //ecommerce. cs. wustl. edu/emediator/ z E-commerce server. z Driven by mobile agent technology. z Performs a variety of e-commerce services. y. Has a configurable auction component that supports a variety of generalized combinatorial auctions, price setting mechanism, novel bid types. y. Has a leveled commitment contract optimizer that determines the optimal contract price and decommitting penalties. y. Has an exchange planner that enables unenforced anonymous exchanges by dividing the exchange into chunks and sequencing them to be delivered safely in alternation to buyer and seller. 54

e. Mediator: Services ze. Auction. House y. Free-to-use third party auction site; y. Wide range of customizable auction types; y. Allows you to buy and sell items as well as to set up markets; y. Allows bidding on combinations of items, and uses novel efficient algorithms for determining winners in this setting; y Allows bidding with price-quantity graphs; y. Supports easy creation and safe on-site execution of mobile Java agents that can buy and sell goods on the server, and set up auctions; y. Implemented in mostly in Java with some computationally expensive matching algorithms being written in C++. 55

Optimal Winner Determination z In a combinatorial auction individual bids by the same agent are joined by non-exclusive OR. z Problem: decide which bids win as to maximize the sum of the bid prices. z Cannot be done in polynomial time (unless P=NP). z Approximate winner determination does not provide worst-case guarantee in polynomial time: y No polytime algorithm can guarantee an allocation within 1/(n 1 -e) bound from optimum for any e>0. z Uses highly optimized tree-search-based heuristics; y Exponential worst case; y Scales up well in practice. z Simplification possible through enforcing special bid structure: y XOR bids; y Stipulating maximal number of combinations to be excepted. 56

Price-Quantity Graphs z Bidders can express continuous preferences, e. g. when a bidder buys a large quantity it will only take lower price per unit. z Curves are piecewise linear for convenience of winner determination. z In a single-sided auction with PQG bidding the winner determination proceeds as follows y. Sum the demand for every unit price; y. Pick the aggregate solution that maximizes the unit price under the constraint that not more is demanded than is available; y. Each bidder gets the amount that it bid at that unit price. z Extensions to double auctions. 57

Mobile Agents z User have agents participate in auctions on their behalf while their computers are offline. z Agents execute on the agent dock which is on (or near) the host machine of the e-commerce server; y. Gives mobile agents safe execution platform to x. Bid; x. Set up auctions; x. Travel to other auction sites; x. Observe activity at various auctions. y. Reduces network latency; y. Key issue in time-critical bidding. z Uses commercial Mitsubishi’s Concordia agent dock: http: //www. meitca. com/HSL/Projects/Concordia 58

Concordia z CONCORDIA is a framework for development and management of network-efficient mobile agent applications for accessing information on a device supporting Java. z Concordia applications: y. Process data at the data source; y. Process data even if the user is disconnected from the network; y. Access and deliver information across multiple networks (LANs, Intranets and Internet); y. Use wire-line or wireless communication; y. Support multiple client devices, such as Desktop Computers, PDAs, Notebook Computers, and Smart Phones 59

HTML Interface z. Users instruct agents. z. Automatic generation of Java code for mobile agents before launching; y. Allows non-programmers to create and launch their agents. 60

e. Mediator: Services ze. Committer, Leveled commitment contract optimizer y. Improves the efficiency of contracts by allowing decommitting; y. Optimizes the contract price and the decommitment penalties for all contract parties based on a game-theoretic model; y. Determines exactly when each party should decommit. 61

Contract Management z Usually a contract is binding y Can’t undo old commitments to accommodate new events. z Anticipated and true development y. An agent accrues information with time; y. This may change his perception of the profitability of the contract; x. E. g. , tasks more costly than anticipated; x. New offers more lucrative. z Leveled commitment contracting protocol serves to alleviate the above: z Agents accommodate future events y. Option of unilateral decommit; y. Decommitment penalty. 62

e. Mediator: Services (cont’d) z e. Exchange. House, a safe exchange planner y Helps avoid non-delivery in exchanges; y Game-theoretic method for guaranteeing that each party is motivated to carry through with the exchange instead of vanishing; y Algorithms for chunking the exchange into parts; y Algorithms for sequencing the chunks to achieve safe exchange. z Coalition formation support (coming soon) y Electronic meeting place and discussion forum y Efficient algorithms for coalition structure generation, and for dividing the coalition's payoffs 63

e. Mediator: Services (cont’d) z e. Voter (coming soon) y Nonmanipulable third party voting protocol; y Game-theoretically guarantees that voters are motivated to vote truthfully. z Meta-auction (coming later) y For finding out what is for sale on the Web. z Reputation databases and algorithms (coming later) z Collaborative rating of goods (coming later) 64

Future Directions z. Agents help buyer & seller y. Combat info overload; y. Expedite specific stages. z. First generation agents y. Create new markets; y. Reduce transaction costs. 65

Future Directions (cont’d) z. Agent technologies need to better manage y. Ambiguous content; y. Personalized preferences; y. Complex goals; y. Changing environments; y. Disconnected parties. 66

Future Directions (cont’d) z. Standards y. Unambiguously & universally define x. Goods & services; x. Consumer & merchant profiles; x. Value-added services; x. Secure payment mechanisms; x. Interbusiness electronic forms. 67

Future Directions (cont’d) z. Further y. New types of transactions; x. Dynamic relationships among unknown parties. y. Create dynamic business partnerships that exist only as long as necessary. 68

Future Directions (cont’d) z 3 rd generation agent-mediated ecommerce. z. Markets with perfect efficiency. 69