Вы находитесь на странице: 1из 22

Problem Restructuring in Negotiation Author(s): Katia P. Sycara Source: Management Science, Vol. 37, No.

10, Focussed Issue on Group Decision and Negotiation (Oct., 1991), pp. 1248-1268 Published by: INFORMS Stable URL: http://www.jstor.org/stable/2632399 Accessed: 23/12/2008 02:23
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=informs. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.

http://www.jstor.org

MANAGEMENT SCIENCE Vol. 37, No. 10, October 1991 Printed itn U.S.A.

PROBLEM RESTRUCTURING IN NEGOTIATION*


K ATIA P. SYCARA Pennsy)lvania15213 Science, Carnegie Mellon University, Pittsbursgh, School of Compupter
To achieve movement towards a negotiated settlement, it is often necessary to restructurethe problem under negotiation. Problem restructuringcan lead to changed perception of the issues by the parties, thus breaking deadlocks and increasing the parties' willingness to compromise. We present a framework and mechanisms for problem restructuringbased on the goals and goal relationshipsof the negotiatingpartiesas well as means of manipulatingthe parties'utility estimates. task. The In addition, previous negotiations are a source of heuristic advice in the restrtucturing restructuringapproach has been implemented in the PERSUADER, a computer program that acts as a labor mediator in labor management disputes. To achieve its task, the PERSUADER negotiates separatelywith each party, company and union, to guide them in reaching agreement. (NEGOTIATION; PROBLEM RESTRUCTURING; PERSUASIVE ARGUMENTATION; CASE-BASED REASONING; AGENT MODELING; BELIEF REPRESENTATION)

1. Introduction During negotiations, participants must shift their positions (e.g., make concessions) in order to reach an agreement. Decision theoretic approaches (e.g., Shakun 1988, Gupta 1989) formulate concession making as a sequential decision process where the next decision of a party depends on the current state, the magnitude of concessions already made, and anticipated responses of the other party(ies). The structure of a negotiation problem can be expressed in terms of variables denoting negotiation goals and issues, relations and constraints among the variables and reservation prices that denote the minimum acceptable levels at which constraints can be satisfied.' The negotiation process is an iterativesearch for appropriatechanges in the goals and constraintsof the participants in order to achieve resolutions that are members of a changing joint solution set (which initially may be empty). The final agreed upon solution is a settlement with values for the negotiation issues, so that the participants'goals are satisfied at some aspiration level, and so that constraints are not violated. It is a fact of life in negotiations that deadlocks occur. One way to remove deadlocks the problem. Problem restructuringis the process of dynamically changing is to restructutre the structure of the negotiation problem to achieve movement towards agreement. In general, some type of restructuringis necessary when one of the parties refuses to accept a proposed settlement. Although the whole process of concession making can be viewed assigned as a (trivial) type of restructuring,in the sense that a concession changes the valule to a variable, in this paper we reserve the use of the terms "restructuring"and "reformulation" for changes in the variables themselves that result in changes in the structure of the negotiation problem. This paper focuses on the restructuring processes that are implemented in the PERSUADER computer program (Sycara 1987). The PERSUADER simulates hypothetical labor-managementnegotiations. The system integratesArtificialIntelligence(Al) concepts
* Accepted by Kalyan Chatterjee, Gregory Kersten and Melvin F. Shakun, acting as Special Editors. This paper was received February 1990; revision received October 1990. ' In the literature (e.g., Shakun 1988, Keeney 1976) a distinction has been drawn between issues, called operationalgoals, and beliefs and desires, called nonoperationalgoals. Since both goal categories may be subjects of restructuring,in the PERSUADER, they are uniformly structured in terms of goailgraphls (see ?4.2) and are uniformly treated by the restructuringprocesses. Accordingly, we adopt the word "goal" to refer to negotiation issues as well as beliefs and desires of the parties. 1248 0025-1909/91/37 10/1248$01.25
Copyright (C 1991, The Institute of Management Sciences

PROBLEM RESTRUCTURING IN NEGOTIATION

1249

and techniques, such as frame-basedrepresentationsand case-basedreasoning (Kolodner, Simpson and Sycara 1985) with graph search and use of multi-attribute utilities (Sycara 1988) to come up with a methodology and mechanisms to enable intelligent machine agents to automatically compute and propose problem restructuringsduring simulated negotiations. Automated problem restructuring is in general a computationally explosive process, given the great number of constraints,and goals present in any given negotiation problem. Therefore, exhaustive search is in general impractical. This necessitates the incorporation in the system of mechanisms to focuts the attention of the problem solver on promising reformulations. Another characteristic that compounds the difficulty of restructuringis the fact that restructuring is a context-sensitive process that, for a given negotiation problem, also depends on ( 1 ) the beliefs and attitudes of the negotiating participants, (2) the negotiation history, and (3) the negotiation context. Hence, the restructuring methodology should incorporate explicit mechanisms that link these factors to the restructuring heuristics. In the PERSUADER, these factors are modeled and represented in object-oriented semantic networks that include ISA hierarchies and inheritance. Each concept is representedin terms of a set of attributeswhose values can be numeiic, symbolic or another concept. The PERSUADER engages in the following types of problem restructuring: (1) introduction of new goals, (2) goal substitution, (3) goal abandonment, and (4) changing the reservation prices of the negotiating parties. In order to be able to deal with the complexities of restructuring,the PERSUADER employs the following restructuringmethods: 1. Case-Based Reasoning (CBR), which consists of retrieving from memory and adapting previous compromises of similar disputants (Kolodner, Simpson and Sycara 1985, Sycara 1987). Because negotiation histories and salient information (see ?3) of the negotiating process through which the parties reached agreement is included in each past case, a reasoner has access to past reformulations as well as strategies for adapting previous cases to produce promising reformulations for the current case. 2. Situation Assessment which representsand recognizes negotiation problems in terms of their abstract causal structure. The representational vehicles that are used in situation assessment are called Situational Assessment Packets (SAPs). They embody the causal knowledge and provide domain independent problem restructuring strategies (Sycara 1987). 3. Search of agents' goal graphs to determine interrelations amongst goals of an agent (Sycara 1987). Since the parties view the negotiation subjectively in terms of their goals and desires, by having access to information concerning goals and relations among them, a reasoner can produce promising reformulations. 4. Persuasiveargumentation,and in particulargeneratingthreats and promises (Sycara 1990a). The effect of arguments is to change the beliefs and behavior of the negotiating parties. In particular,the goal graphs of the parties are changed. This restructuringresults in changing the perception of the parties concerning the negotiation. Case-basedreasoning can be used for all types of problem restructuring.This is possible since each previous case includes information about negotiation impasses and the means by which they were overcome, i.e. the repairs made to rejected proposals so as to achieve movement towards agreement. Problem restructuringis one type of repair strategy. Associated with each repair of problem restructuringtype is also the restructuringcategory, namely goal introduction, goal substitution, goal abandonment and changing of reservation prices. Thus, when a problem solver accesses similar previous impasses, it can select out of the previous impasses one(s) whose repairs seem most appropriate in the currentsituation. In contrastto CBR that accesses and adapts previously used restructuring strategies, situation assessment, goal graph search and generation of threats and promises are generaltivemethods.In other words, they generate reformulations from scratch.

1250

KATIA P. SYCARA

?2 gives a brief summary of the PERSUADER system, presents requirements for problem restructuringand the overall restructuringprocess; ?3 presents an overview of case?4 presentsthe generativerestructuring based reasoning and how it is used in restructuring; processes; ?5 presents the generative restructuringprocesses to achieve goal introduction; ?6 presents the generative algorithms for goal substitution; ?7 presents the generative algorithms for goal abandonment and ?8 generative changing of reservation prices. Concluding remarks are presented in ?9. 2. The PERSUADER System Our model of problem restructuring is part of a general multi-agent, multi-issue negotiation model (Sycara 1987, 1990b) that has been implemented in a computer program, the PERSUADER. In contrast to knowledge-based work on negotiations that has concentrated on providing support for human negotiators (Kersten et al. 1990, Jarke 1987, Goeltner 1987), our work concentrates on automating the dynamics of the process itself and coming up with a settlement that the parties agree on. Although the PERSUADER can operate autonomously, the interface allows users playing the role of union and company to register reactions to the system's suggestions and provide feedback. The PERSUADER system involves three agents: a company, its trade union and the mediator whose task is to help the other two agents reach an acceptable compromise. The mediator is engaged in parallel negotiations with the union and company agents. The PERSUADER's input is the set of conflicting goals of the company and the union, and the negotiation context. The negotiation context consists of factors, such as the general state of the economy, the state of the industry to which the company belongs, the financial situation of the company, and the profile of the international union to which the local union belongs. The final output is either an agreed upon settlement (contract) or an indication of failure if the negotiating parties did not reach agreement within a particular number of proposals (to simulate the inability of parties in the real world to reach agreement before a strike deadline). A contract consists of a set of issues, such as wages, pensions, provisions for seniority, vacations, etc. and a set of values assigned to the issues. The final agreed upon contract is the result of iterative contract modification by the parties and the mediator. The PERSUADER uses two categories of knowledge to perform its tasks: (a) domain knowledge and (b) reasoning knowledge. Domain knowledge includes knowledge about negotiations, negotiators, negotiators' goals, negotiation context and negotiation settlements. Each concept is represented in an object-oriented manner using a network of frames and the inheritance mechanism (Sycara 1987). Each frame has slots that represent salient features of the represented concept. The filler of each slot is either a value or another frame representing an appropriate concept. For example, the concept of a local union is represented as a frame that has slots to represent the name of the union (the slot filler is a string denoting the union name), the union leader (the slot filler is a frame network representing attributes of the concept of a union leader), the union goals (the slot filler is a goal graph represented as frame network), the international union to which this local union belongs (the slot filler is a frame network), the bargaining power of the local union (the slot filleris a procedurethat takes as argumentsattributesof the economic context and the international union, and returns numerical values), the composition of the bargaining unit (the slot filler is a frame network representing the bargaining unit in terms of job classification, number of skilled, unskilled and semiskilled workers, age categories and degrees of militancy), the time the local has been in the company (the slot filler is a number whose units are months), and the relations of the local with the international (the slot filler is a string that takes the value "good" or "bad"). For more details on the representation, see Sycara (1987). Reasoning knowledge includes the knowledge needed to assess the "fairness"of a solution, and knowledge needed to improve a rejected solution. This knowledge is expressed in terms of multi-attribute utilities as-

PROBLEM RESTRUCTURING IN NEGOTIATION

1251

sociated with the parties' goals, and criteria for selecting negotiation proposals (Sycara 1987). To use the system in a new domain, cases should be structured in terms of the salient features in the new domain. The reasoning algorithms are domain-independent. Some of the reasoning knowledge that is used in our model has its conceptual origins in standards that are commonly used to justify arbitration awards (Elkouri and Elkouri 1972). The standard that is most frequently used is the prevailing practice standard, namely the bargainingbehavior of similar parties.In our work, this standardgets abstracted to reasoningfrom precedentcases. This is a suitable method for the domain since previous cases represent good solutions to the difficulties that are endemic to finding acceptable compromises in multi-agent negotiations. Knowledge acquisition in the PERSUADER is based on four main sources: ( 1 ) the labor relations literatureincluding books, journals, newspapers and marazines, (2) published arbitration awards, where the arbitrator cites the facts and criteria used in the decision, and (3) two human expert mediators,2 and (4) the PERSUADER's Case-Based Reasoning process. Nine months were spent tapping these sources and making decisions on the knowledge to be initially incorporated in the Case Knowledge Base. The system is started with a set of cases that are placed in its Knowledge Base. At the end of each problem solving session the PERSUADER's memory is updated with information from the newly resolved case as well as any new information that has been acquired during problem solving (e.g., from the parties' feedback). Thus, knowledge acquisition and learning is ongoing since it is a by-product of problem solving. The negotiation process consists of iteration and interleaving of three main tasks: generation of a proposal, generation of a counterproposal based on feedback from a dissenting party,and persuasiveargumentation.These tasks can be iteratedand interleaved several times during negotiation. The mediator agent in the PERSUADER system generates an initial compromise proposal and presents it to both the union and company agents who evaluate the proposal from their perspectives and give the mediator their reaction. If both accept the proposal, then it is the final compromise. If one of the agents rejects it, the mediator enters a negotiation with that party and makes a decision whether to change the proposal, or attempt to change the disagreeing party's position. Restructuring takes place as a result of an explicit rejection by one of the parties and may result in changing all concerned parties' perceptions regarding the negotiation. A successful problem restructuringshould at least change the perception of the party whose disagreement hinders progress towards resolution. The desired outcome of problem restructuringis a new proposal that gives the rejectingparty a higher payoff than the rejected proposalbut does not lower the payoff of the partythat had accepted the previous proposal to such an extent that it will reject the new proposal. 2.1. Requtirements for Problem Restructutring The restructuring task is knowledge- and search-intensive. Sources of restructuring knowledge can be found in the abstract causal structure of a negotiation problem (see ?4.1 ), in relationships among the negotiating parties, in beliefs and desires of the negotiators and in relationships between the negotiation and the environment. In order to automate restructuring,this knowledge must be captured and represented in knowledge structures explicitly and in a principled way. Unlike rule-based systems that use the rule formalism as a sole representational vehicle, the PERSUADER integrates a variety of representational mechanisms for knowledge structuring and problem restructuring: ( I ) Situational Assessment Packets (SAPs) that capture the abstract causal structure of a negotiation problem and store general and specific strategies for restructuringand resolution, (2) Negotiation Cases that capture specific instances of negotiations, (3) Goal Graphs that represent goals/issues, as well as beliefs and desires of the participants, (4)
2 One is a Professor of Economics and the other came up through the ranks of the International Machinists' Union.

1252

KATIA P. SYCARA

Utilities associated with the issues under negotiation and with higher level goals of the parties, and (5) Rules that facilitate inferences. This representationalsynergy is necessary to enable the system to reason at a deeper level of understanding. We have identified some characteristicsof the task that give rise to a set of requirements for automated problem restructuring. * The number of variables that could enter the restructuringprocess is very large and not easy to identify in general. In order that a reasoner produce an effective formulation, it needs to be able to represent and maintain sets of possible goals of the agents it is interacting with. * Because the set of goals of an agent could be large, and because different types of reformulation are appropriate under different circumstances, a reasoner must be able to focuis its attention on promising reformulations in terms of (a) type of restructuring (change of reservation prices, goal abandonment, goal substitution, or introduction of a new goal), and (b) the specific goal(s) that must be introduced or abandoned. * A particular restructuringmay or may not be able to resolve a deadlock. Based on the parties' feedback, a reasoner must be able to update its knowledge and inkferences concerning the goals of the parties. The input to the restructuringprocess, depicted in Figure 2-1, is a rejected proposal3 (the contract issues and their values), the rejected issue and the reason for rejection provided by the rejecting party. These provide focal points for the PERSUADER, from which search for promising reformulations will be conducted. We make the assumption that a party rejects a proposal if it does not give it high enough payoff. So, a reformulator should look for reformulations that result in new proposals with increased payoff for the rejectingparty. On the other hand, the proposal that resulted from a reformulation should not inordinately decrease the payoff of the party that had accepted the previous proposal. The criterion that the PERSUADER uses is that a restructuringshould result in proposals that increase the rejecting party's payoff by a greater amount than they decrease the payoff of the party that had previously accepted the proposal. The first restructuring strategy the PERSUADER uses is to introduce a new goal in the negotiations. The new goal should be such that it compensates for some loss of payoff for the rejecting party without appreciably decreasing the payoff of the accepting party. Introduction of a new goal is tried first because we believe that it is the mildest form of reformulation, not requiring abandonment of parties' goals. Introduction of new goals can be achieved by the restructuringprocesses of case-based reasoning, goal graph search or situation assessment. At each iteration, the PERSUADER attempts to introduce a goal that could result in a promising reformulation. If all the promising reformulations have been exhausted without breaking the deadlock, then goal substitution is attempted. The PERSUADER attempts to substitute a suitable goal for an issue/goal that has been objected to. Goal substitution can be seen as a relaxation of a party's demands that gives the party lower payoff. Goal substitution is accomplished through goal graph search and case based reasoning. If goal substitution fails to resolve the deadlock, the PERSUADER pursues goal abandonment of the objectionable issue/goal. The process through which goal abandonment is achieved is the generation of threats and promises, or case-based reasoning where a previously used threat/ promise is retrieved and adapted to the current situation. If goal abandonment also fails to resolve the deadlock, then the PERSUADER pursues change in the reservation price of the rejecting party. This is accomplished by generating threats to break off negotiations, or retrieving appropriate negotiation break off threats through case-based reasoning. Since case-based reasoning can be used for every category of problem restructuring, we devote the next section to a brief overview of case-based reasoning and description

3See ?3 for a description of case-based generation of proposals.

PROBLEM RESTRUCTURING IN NEGOTIATION

1253

e ecte Issue and Reason

Goal Introduction
t

_
\

Goal Graph

<_ ~~Search
Situation Assessment

Stop Restructuring

les s"Ye

eadlocX olved

Goal Substi tution

~
rap

Stop ig

_ ea yp eol Restructuring

<

Abandoment Overview ofCae-BasPersuasive / \ Stop ia ar s emsloc ess bae e Ps ae Restructuoeing 3. Change of Reservatdion i t Price ta Argum entat ~~i on ore-usefpaisesol

red t have art

|Restructur ing

So
NJ Nn

Restructuring FAILURE
FIGURE 2-1.

The Problem Restructuring Process.

of the process by which repairs are accessed and the various types of restructuringperformed. 3. Overview of Case-Based Reasoning

Case-based reasoning is a problem solving process based on re-use of past solutions to

address a new problem. The PERSUADER's case memory contains successes, i.e. cases that resulted in mutually accepted settlements in past similar circumstances. They are used as basis in resolving similar disputes. The PERSUADER also stores failures, i.e.,
.cases where no settlement has been agreed upon in the allotted time, and the reason for the failure. Failures can be recalled in situations that have similar features to the one where the failure occurred. Previous similar failures warn a problem solver about potential diffinculties that might arise in the current case. The negotiation history for a case is

1254

KATIA P. SYCARA

jection), the rejection reason (if one is offered) and the PERSUADER's repairsto improve a rejected proposal. Successes are indexed under salient domain features, such as industiy to which the disputant company belongs, geographical location of the company, job classification of the employees etc. Impasses have three additional indices, an index indicating "failure,"the failure cause, and the negotiation issue(s) that was (were) involved. As the PERSUADER settles an increasingnumber of negotiations, its case base is enriched, making available a great variety of previous cases it can reason from. Cases are organized in memory hierarchicallyaround important concepts in the problem domain. In order to perform Case-Based Reasoning, cases need to be retrieved in terms of conceptual similarity. The basic idea behind conceptual similarity between two concepts is that they have important common attributes. For example two truck companies are similar because they share the common attributes of being a company and having transportation of goods as their product. Similar concepts are organized into larger groupings based on their similarities, and differentiated from each other in terms of their differences. The two truck companies of the example are organized under the more general concept of "ground transportation companies" and differentiated by such differences as location, size, financial condition etc. The high level knowledge structure that we use to organize similar concepts in memory is called a generalized episode (Kolodner 1984). Generalized episodes organize cases into a hierarchical network whose nodes are either another generalized episode or an individual case. The PERSUADER uses CBR to perform a variety of tasks. For each of these tasks, a separate part of a case is used to guide the reasoning. For example, in generating an initial compromise, the system reasons from previous settlements;in generatingpersuasive arguments, it uses previous arguments; in performing problem restructuring,it uses previous negotiation impasses. Although different knowledge is used in each of these situations, the Case-Based Reasoning algorithm is the same. The steps of the algorithm are as follows: ( 1) Retrieve similar cases from the Case Memory,4 (2) Determine the case that is most similar to the current negotiation (the best case), (3) Extract knowledge from the best case, (4) Construct a baseline solution (called the "ballpark" solution), (5) Adapt the baseline solution to fit current circumstances. At the end of problem solving, memory is updated with the new, successful or failed, case. To generate an initial settlement, knowledge is extracted from the contract of the most similar precedent case and adjusted through standard adjustments to form the baseline solution. After further checking and adaptation to fit as closely as possible the characteristics of the current negotiation, the adapted solution is proposed to the parties. For more details of the use of CBR in generating proposals, see Sycara ( 1990b). 3.1. Problem Restructuring Using CBR Each objectionable issue and reason that the rejecting party offers are used as indices in the case memory to select impasses with the same stated impasse goal and impasse cause as in the present deadlock. Examples of reasons for rejection are inability on the part of the company to pay for a proposed economic demand, consideration of a concession as too low by the union, consideration by the company of a proposed strict seniority clause as restriction of its management rights. The selected impasse supplies a repairthat can facilitate impasse resolution. We have identified several general categories of repairs, such as changing the value of an issue, restructuring the problem, passing costs of a settlement to third parties (e.g., the consumer), invoking authority, extending the current agreement and resorting to binding arbitration. These repair types are organized in abstraction hierarchies in a frame-based representation. Figure 3-1 depicts part of such an abstraction hierarchy.
4By "cases" we mean appropriatecases or parts of cases, such as a previous argument.

PROBLEM RESTRUCTURING IN NEGOTIATION Repair Is-a Is-a is-a

1255

Problem s-a

Restructurin sa

Extend Agreement s-a Pass-cost-consumer i on

Pass-cost-third-party s-a \Is-a Gov t-subsidy

Goal-substitution

Goal-abandoment Change-reservat

Goa 1- i ntroduct in
FIGURE

3-1. Partial Abstraction Hierarchy of Repairs.

Goal introduction, goal substitution, goal abandonment and changing of reservation prices are refinements of the repair of type "problem restructuring".Information about the various repair types and their refinements is kept as part of the repairs associated with cases in memory. The PERSUADER can check and select suitable repair types at the appropriate time during problem solving. The process of problem restructuringusing case-based reasoning is as follows:5 1. Construct a memory probe consisting of the conjunction of indication "FAILURE", the rejecting party, the rejected issue and cause for rejection. 2. Using the probe retrieve similar failures (impasses). 3. Use additional similarity criteria, such as industry to which the company belongs, geographical location, job classifications, dispute context etc. to select the most similar set of impasses to reason from. 4. Out of the set of impasses identified in step 3, select those whose repairsare of type "problem restructuring".Access the repairs. 5. Out of the repairs identified in step 4, do: o Select the repairs that are of type "X".6 * Rank order the selected repairs in terms of applicability to the current case. * Successively propose the repairsto the rejecting party in an attempt to resolve the impasse. If one of the proposed reformulations succeeds in breaking the deadlock, then stop. * If the reformulations of type "X" have been exhausted without success, use generative approaches (see ?4) to construct additional reformulations of type "X". As an example, consider the PERSUADER during deadlocked negotiations of Porter Structural Inc., an airframe manufacturer. One of the important union demands is for job security. To operationalize this goal, the union had demanded that the company should not be allowed to subcontract. The company refuses saying that a no-subcontract clause restricts its flexibility in decision making. The union insists saying job security is very important. Deadlock occurs. The case memory is searched for suitable repairs. Searching memory with index FAILURE, UNION, SUBCONTRACT, JOBSECURITY 7 impasses found Select impasses with repair type "restructure"
We present the algorithm for the type "goal introduction". The algorithm is the same for all types of restructuring. 6 Depending on the point in the restructuringprocess (see Figure 2-1 ) the system is in, the variable X gets instantiated to "goal-introduction","goal-substitution","goal-abandonment",or "changing-reservation-prices". The algorithm is the same for all types of restructuring.

1256

KATIA P. SYCARA

applying features similar industry, same area, same job classification 3 impasses found Select impasses with restructuretype "goal introduction" 1 impasse found (impasse3) Looking at repair "The company should pay for worker retraining.. ." from impasse3 Since retraining costs low relative to company profits repair seems applicable Propose that retraining goal be introduced Along with the repairs,applicability conditions of the repairsare stored. An additional check is made to determine whether the company can afford retraining costs. Retraining costs are computed for a maximum of 3 of the bargaining unit members for the next three years. Costs are compared to the company's projected profits for the next three years. Costs are found to be 3% of projected company profits, so retraining is proposed. The introduction of retraining breaks the deadlock since worker retraining is an instrumental goal to the higher level goal of job security and does not restrict the company's decision-making flexibility. After further discussion, a lower maximum number of retrained workers (4 instead of 3 ) is agreed upon and some limitations are also placed on the current subcontracting language. In other words, once an acceptable restructuring has taken place, negotiation and concession making may continue for agreement to be reached. 4. Overview of Generative Restructuring Processes 4.1. SitutationAssessment Situation Assessment is the process through which the PERSUADER handles problem in atypical situations. In such situations, restructuring restructuring is even more important since strategies used in more typical negotiations will probably not work. In order to address typical negotiation situations, an automated reasoner must have knowledge that allows it to make inferences that result in expectations about the impact of its decisions. In negotiations, we have identified the following sources of expectations: (a) prevailing behavior of similar agents (prevailing practice), (b) beliefs about the rationality of the agents, (c) beliefs about the temporal continuation of a state, and (d) roles and relationships among the disputants. Departures from typicality, captured in knowledge structures called Situational Assessment Packets (SAPs), take the form of expectation violations associated with the above knowledge sources. SAPs that we have identified and implemented in the PERSUADER include SAP MISMANAGEMENT, where one negotiating party mismanages a resource necessary for the achievement of goals that are common to other parties in the negotiation; SAP IDEOLOGY where a party refuses some tangible benefits for ideological reasons; SAP MISPLACED-LOYALTY where a party is pressuredto forego some benefits to itself out of loyalty to another etc. (for more details on SAPs, see Sycara 1987). SAPs capture knowledge about the abstract causal structure (see Figure 4-2) of negotiations and relate it to violation of expectations. Being able to access causal knowledge that it can use to base its inferences on, enables an automated reasoner to focus its attention on goals that produce promising reformulations. One of the important kinds of knowledge that SAPs express is the interrelations among multiple goals, particularly in atypical situations. The novel solutions needed to resolve atypical disputes could not have been found without knowledge of goal interrelations (for a detailed example, see ?5.1). SAPs express abstract causal structures of negotiations that transcend particular domains. In contrast to cases, the features incorporated in a SAP are features that pertain

PROBLEM RESTRUCTURING IN NEGOTIATION

1257

to the negotiation planning process itself rather than to information about particular negotiations. In other words, these features contain maeta-planning information (infor-

mation to the planner of how to plan). This has important consequences. If two situations can be recognized in terms of the same SAP, they can be handled by the same metaplanning strategies. Thus, SAPs enable a problem solver to access general advice about problem solutions and utilize solutions from analogous problem situations in different domains. The ability to store situations from different domains under the same SAP makes SAPs very powerful mechanisms. While there may be no precedent in the labor management domain, for example, that can help a mediator deal with mismanagement problems, general experience with mismanagement in other domains can be helpful in coming up with a solution. Organizing similar memories at the abstract level of goals and plans, as SAPs do, facilitates sharing of information between domains and contexts that makes finding creative solutions possible. Once an episode has been recognized in terms of a SAP, its problem solving heuristics become available for use in situations with different content. Thus, once problem solving advice has been learned in one context, it can help processing in a different context, if the experience was recognized in terms of an appropriate SAP. SAPs contain abstract structures that represent: ( 1) a problem solving situation, (2) expectations associated with the situation, (3) the reason expectations are violated, (4) who/what is responsible for the violation, (5) strategies and plans to find an acceptable solution, and (6) how to justify the solution. Figure 4-1 presents the conceptual structure of the SAP MISMANAGEMENT. To see how problem restructuringis done, we must first examine how various pieces of knowledge and inferences are linked inside a SAP. The causal structure of a SAP is a
graph whose nodes represent goals, states and actions of the agents. The nodes of the

graph are connected via several types of links. A link between two goals denotes whether one goal is instrumental to the other; a link between an action and a goal (or a state) denotes whether the action is a precondition of the goal (or results in the state); a link between two states denotes the relationship between the states; a link between a state and a goal denotes the causal relationship between the goal and the state (see Figure 4-2). The overall organizationalstructureinside a SAP is hierarchicalfrom general to specific knowledge (see Figure 4-2). A hierarchical organization of knowledge within a SAP allows a problem solver to access general strategiesfirstand, depending on the information available, subsequently access more specific strategies and plans. Thus, the general strategies are at the top of a SAP. Under general strategies more specific strategies are stored. Under those, applicable plans are stored. Once a situation has been recognized as fitting

SAP MISMANAGEMENT recognition criteria: 1. x and v have a noncompetitive high level goal G7 2. x mismanages some resource, the possession of which is an enablement condition C for the achievement of G. 3. G is in danger of failing solution: an equitable solution to prevent the failure of G is to have x, the guilty party, bear the brunt of the recovery cost by denying him a prerogative or a reward justification: appeal to theme of fairness and add that if Y does not perceive the solution as just, then v? will not cooperate and thus G will fail (which x certainly does not want).
FIGURE

4-1. Content of SAP MISMANAGEMENT.

7This might seem to be a contradiction in the case of companies and unions. It is not so, however, since the highest level goal of both company and union is to produce goods for which both get rewarded.

1258

KATIA P. SYCARA Higher level goal GO(shared by agents x and y)

Instrumental goal

Instrumental goal

goal G1 contributes to GOsuccess (possessed by agent [xl)

goal G2 contributes to GOsuccess (possessed by agent [y])

cause negation of G1

Mismanagement action [x]

goal G3 Fairness goa

Interfered by failing GO

plan to achieve goal

Condition: Guilty party known Strategy: Punish only guilty party

*
plan specialization Condition : Party has prerogative Plan: Deny prerogative
FIGURE 4-2.

Structure of SAP MISMANAGEMENT.

the structure of a particular SAP, the SAP is accessed. The internal causal structure of a SAP then guides the problem solver in proposing a promising reformulation. 4.2. Goal Graph Search Goal graphs are directed acyclic graphs that are used as the representational vehicles for modeling parties' goals and beliefs. These models are used as sources of restructuring advice and as consistency checkers. As will be described in detail in ?5.2 and ?6, goal graph search is used for goal introduction and goal substitution.8 Goal graphs are not static structures. A party's goal graph is updated based on the reaction to problem restructuring. In this way, during the course of a negotiation an agent's model is refined and corrected. This functionality is important in automating negotiations since (a) it is not possible for an agent to have an a priori correct and detailed goal graph of another,
8

The generation of threats and promises also uses goal graph search.

PROBLEM RESTRUCTURING IN NEGOTIATION PROFITS (+) PRODUCTION-COST (-)


M-ATERIALS COST MATERIALS-COST

1259

SALES (+) / QUALITY (+) PRICES (-)

PLANT-EFFICIENCY( EMPLOYEE-SATISFACTION (+) ECONOMIC-CONCESSIONS

LABOR-COST (-)

\
EMPLOYMENT (-) ECONOMIC-CONCESSIONS (-)

T WAGES (

NON-ECONOMICCONCESSIONS (+)

4
)

WAGES (-)

FRINGES (-)

AUTOMATION

SUBCONTRACT (+)
FIGURE 4-3.

A Company's Partial Belief Structure.

and (b) goals are not static but change with external circumstances and an agent's experiences. An example of a goal graph is shown in Figure 4-3.9 Each node in a goal graph representsone of the agent's goals. Arcs in the graph linking two goals represent the relationship between goals in terms of how one affects positively or negatively the achievement of the other. Associated with each node is: * a sign (+ or -) that denotes the desirability of an increase or decrease in that goal. For example, PROFITS( +) represents the company's goal of increasing its profits. * the amnount by which the goal should be increased or decreased. The amount's units can be either a percentage, an absolute dollar figure, a range of values, or a symbol, such as "a lot". The units chosen depend on the requirements of the domain. We use percentages and dollar units and the associated meaning of "at least", or "at most" the denoted amount. The value 0% is used to represent preservation of a goal. For example, FRINGES(-, A = 5%) indicates that the company desires to decrease the cost of fringe benefits by at least 5%. * the imnportance that the party attaches to this goal. For example, increasing profits PROFITS(+, A = 8%, I - 10) means that the company's goal of increasing its profits by at least 8% has an importance of 10. The importance vaiiable assumes values from 0 (least importance) to 10 (greatest importance). * the feasibility, as perceived by the party, of achieving this goal. For example, the feasibility of achieving an increase in sales SALES(+, A = 6%, I = 9, F = 0.7) has a feasibility of 0.7. Feasibility is a probability measure and takes values from 0 to 1. The nodes at the top of a graph denote an agent's highest level goals. For example, PROFITS(+, A = 8%, I = 10, F = 0.6), representing the company's desire to increase its profits by at least 8%,is the highest level goal of the company. The nodes at successively lower levels denote the subgoals through which the higher level goals are achieved. Values associated with each of the edges connecting subgoals to goals denote an assessment about the percent contribuition of the subgoal to the accomplishment of the corresponding goal. The direction of the edges is from subgoals to the goals to whose achievement they contribute. Referring back to the highest node in the figure, we see that profits can be raised, (PROFITS(+, A = 8%,1 = 10, F = 0.6)) by decreasing production costs (PRO9For ease of viewing we draw the goal graphs as trees with only the sign parameter shown.

1260

KATIA P. SYCARA

DUCTION-COST (-)) or by increasing sales (SALES ( +)). The contribution of a sales increase to raising profits is 40% and of reducing production costs 60%. The graph is obviously acyclic since it would be absurd to say that a goal contributes to its own achievement through a series of subgoals. Contribution ranges from - 100% to 100%.A positive value means that the subgoal supports the achievement of the higher level goal. Goal graphsrepresentinterrelationships among goals in a more detailed fashion, thus subsuming the functionality of AND/OR graphs. Conjunction is inferred when the sum of the contributions of subgoals is 100%. Disjunction is inferred when each subgoal's contribution to the same higher level goal is 100%. Allowing negative contributions expands the expressive power of goal graphs. A negative contribution value has the interpretation that the subgoal not only does not contribute to the higher level goal but it is detrimental to it. For example, a negative public image would have a detrimental effect on a company's sales. A contribution value of zero means that the subgoal is irrelevant to the achievement of the higher level goal. A path from X to Y in a goal graph constitutes a causal chain that provides an explanation of the change in Y in terms of the change in X, assuming no other change has occurred in the rest of the graph. The path WAGES(-) to PRODUCTION-COST(-) in the company's goal graph can be interpreted as follows: "Other things being equal, diminishing the cost of wages results in decreasing the cost of the economic concessions, which causes a decrease in labor costs, leading to a decrease in production costs and an increase in profits". In addition to an agent's beliefs, the representationincludes an estimation of his utilities. The concept of utility is the basis for selecting among alternativedecisions. Each alternative is evaluated in terms of a number of attributesthat a decision maker considers important. of an agent and the tradeoffs he is willing to Utilities express the preference structutre make among various attribute values. The utilities of the individual attributes are combined to give the overall utility (payoff) of an alternative. Being able to compare different alternatives enables a decision maker to choose the alternative that affords him maximal payoff. In labor negotiations, the pertinent attributes are the issues under discussion and different contract proposals are the alternative decisions. In the PERSUADER, each party's overall payoff is expressed as a linear combination of the utilities associated with the issues. In order to construct effective reformulations, a reasoner needs to know both the belief structure of a party as well as the influence that the environment would have on the belief structure. In the labor domain, the economic context of a dispute is one of a variety of environmental factors that influence a company's (union's) belief structure. For example, in recession, a union's job security goal has higher importance than in boom. When competition is stiff in an industry, a company's automation goal assumes greater importance, and also has a greater contribution towards the goal of reducing production costs. In the PERSUADER, knowledge of salient features of the environment is recorded and is reflected in the shapes of the utility curves and goal importance in the parties' goal graphs (Sycara 1988). Utility curves are representedby the equations that describe them. 4.3. Persuasive Argumentation Persuasiveargumentationis the processthrough which an agent, the persuacder, attempts to change the beliefs and behavior of another agent, the persueadee.In the PERSUADER, the mediator agent is the persuader and the union and company the persuadees. Our argumentation model has two general methods for generating arguments: (a) construct arguments from scratch, and (b) use case-based reasoning to recall previously used arguments and adapt an appropriate one. In our framework, persuasive arguments are viewed as plans used by the persuader problem solver to accomplish his goal. The goal of a mediator is to help the negotiating parties reach an agreement. This can be facilitated by increasing the parties' willingness to accept a proposed settlement. We claim that a

PROBLEM RESTRUCTURING IN NEGOTIATION

1261

party's willingness to accept a proposal is proportional to its perceived payoff. In other words, a party rejects a proposal because it does not give it enough payoff. Hence, if a persuader could manipulate a party's utilities (resulting in manipulation of the party's payoff), he would be able to affect predictably the behavior of the persuadee. Convincing a persuadee to change his evaluation regarding a proposition is modeled as producing an argument to increase the payoff of the proposition. The issues that appear in a proposition are a subset of the goals that appear in the agent's belief structure. Hence, the task of a persuader can be viewed as finding the most effective argument that will change a persuadee's belief structure in ways that increase his payoff. Since a persuadee's payoff can be approximated by a linear combination of his utilities, his payoff can be increased by ( 1 ) changing the importance the persuadee attaches to a goal/issue, (2) changing the persuadee's perception of an issue's value, and (3) pursuing goal abandonment on the part of the persuadee via threats/ promises. The PERSUADER generates arguments for all three types of desired changes in the beliefs structure of a persuadee (Sycara 1990a). Goal abandonment is one of the ways in which the PERSUADER performs problem restructuring. ?7 presents details of the restructuring algorithm. 5. Introductionof New Goals Movement towards agreement is often facilitated by the introduction of new goals in the negotiation. New goals could provide new contexts within which the negotiation could be perceived. For example, the restructuring of the Southern Airlines negotiation (see ?5.1 ) through the introduction of the concern of mismanagement and blame attribution (via situation assessment) allows for creative resolution of the negotiation. Introduction of new goals could also provide additional ways in which higher-level goals can be achieved, thus compensating for unacceptable aspects of negotiation issues. For example, the introduction of worker retraining in the Porter Structural Inc. negotiations (?3. 1) compensates somewhat for the detrimental effects of subcontracting to job security for the union. In the PERSUADER, besides case-based reasoning, the processes of (1) Situation Assessment, and (2) goal graph search can be used for goal introduction. 5. 1. Goal IntrodulctionThrough Situation Assessmnent 1. During contract negotiations, Southern Airlines presents its employees EXAMPLE with the ultimatum that, if they don't take wage cuts of 8%, the company which has become noncompetitive will go bankrupt. The employees protest and a mediator is called in. The mediator finds out that Southern Airlines has been losing money because of mismanagement in an industry where other airlines are making money. The mediator proposes that the employees accept 5% wage cuts and that the company give stock to the employees as well as accept employee representatives on the board of directors.10 The above example illustrates a situation where the solution to the problem was constructed neither through adaptation of a case-based solution nor by analyzing utilities, but by introducing a novel alternative. Judging the Southern Airlines negotiations as atypical, the mediator came up with a solution neither in the realm of prevailing practice nor predicted by payoff adjustments. To distinguish a typical from an atypical situation, the PERSUADER evaluates the present case vis a vis prevailing practice. This evaluation occurs during applicability checking for knowledge extraction from a selected precedent to construct a ballpark solution. The result of evaluation is an indication of whether the case can be considered typical or not. If so, adaptation of the ballpark solution takes place. If not, Situation Assessment to generate a reformulation of the negotiation problem is activated.
0?Such solutions are not uncommon in real situations, as is attested by the 1985 settlement of EasternAirlines with the International Pilots' Association.

1262

KATIA P. SYCARA

To evaluate case typicality in the above example, the PERSUADER checks that the air transport industry average for wages is 5%increase. Asking for a wage cut of 8%,the company is far outside the range indicated by prevailing practice. It concludes that the case is atypical and activates the process of Situation Assessment. Situation Assessment classifiesatypical situations in categoriescharacterizedin terms of the expectation violation they embody. When an expectation violation stemming from one of the inferentialsources (prevailing behavior of similar agents, role themes, economic rationality, and temporal continuation of a state) is encountered, the process of SAP recognition is activated. Example 1 is processed by recognizing that the Southern Airlines negotiation fits SAP MISMANAGEMENT. SAP MISMANAGEMENT is an example of a SAP that is recognized when expectations arising from prevailing behavior of similar agents are violated. A reasoner would expect a company belonging to a prosperous industry to be making profits. The expectation violation involved in each case controls what parts get focused on. For example, the PERSUADER's knowledge of normality in a company's life allows it to make the inference that bankruptcy is an abnormal situation. Knowledge of what bankruptcy means focuses the attention of the system on the company's finances. What has to be examined next is whether bad financial conditions plague many companies in the industry (as would happen, if, for example, fierce foreign competition caused an industry slump). Once the reasoner finds out that the present company is the only one in the industry incurring heavy losses, then attention gets focused on discovering the behavior of the particular company (or union) that caused the bad financial situation. In Example 1, the noncompetitive goal that the company and union share (cf. Figure 4-1 ) is to prevent the company from going bankrupt, since that would hurt both the company and the union. One characteristic of SAP MISMANAGEMENT is that one of the interacting agents is to blame for the danger of the failure of the common goal. The potential failureis used as the justification for the proposed solution, which would certainly be disagreeableto the guilty party. A problem solver using this SAP appeals to the theme of fairness and "just deserts", which is a new goal introduced in the negotiation. Based on this goal of equity and distributive justice, a whole new range of solution strategies are available. These strategies and the solutions they suggest ( 1) are an answer to the negotiation deadlock, and (2) they could not have been inferred, had the reformulation not taken place. SAP MISMANAGEMENT introduces the high level goal of "fairness"which changes the way the parties view the negotiation situation. The company stops perceiving the situation as one where, because of financial difficulties, it must ask the workers to take severe wage cuts and expect the workersto accept its proposal;since the financialdifficulties of the company are the result of mismanagement, the company must give noneconomic compensation to offset the loss of worker income. The introduction of the "fairness" goal encourages the union to demand and win noneconomic compensations, such as the right to be represented on the board of directors. In general, SAPs contain not only abstract causal structures of negotiations that could be used for restructuring,but also strategies and plans for producing concrete proposals to break deadlocks. As can be seen in Figure 4-2, the "FairnessGoal" has associated with it a set of abstract and specific plans, such as "punish only the guilty party" and "deny prerogative". In the Southern Airlines example, the PERSUADER makes use of the "deny prerogative" plan. Since management rights is a kind of prerogative, the PERSUADER searches the list of management rights for a suitable one (one whose preconditions are satisfied in the current negotiation) to negate. This process results in the suggestion that employee representatives should sit on the board of directors. 5.2. Goal IntrodutctionThrouighGoal Graph Search An advantage of using goal graph search for introduction of new goals is that the traversed graph edges provide explicit linkages among goals. These explicit linkages can

PROBLEM RESTRUCTURING IN NEGOTIATION

1263

be presented to the objecting party to change its problem perception in order to increase its willingness to move towards agreement. Introduction of a new goal is done through matching the objectionable goal in the objecting party's goal graph and traversing the graph upwards (from leaf nodes to their ancestors). Each of these paths shows the consequences that a change in the matched goal will have for higher level goals. The idea is to find as appropriate restructuringcandidates those goals whose contribution to higher level goals is greater or equal to that of the objectionable goal. The algorithm for introducing a new goal through search of the belief structure of the party that has objected to an issue is now presented. Let gi be the objectionable issue causing deadlock. 1. Match11 gi in the goal graph. 2. Traverse the contribution link to gi's parent goal gk.2 3. Traverse the links downwards to the set T of children of gk (siblings of gi). 4. Out of the set identified in step 3, select the goals whose contribution to gk is greater than or equal to the contribution of gi. This set C constitutes the candidates to be introduced.

5. Delete from C all goals that are already present in the current negotiation. 6. Rank order the set C of candidates identified in step 4 in ascending order of contribution to gk-. 7. Successively introduce the elements of set C in the negotiation in ascending order of contribution to gk . If one of the elements of C successfully breaks the deadlock, STOP. 8. If the set of candidates C is unsuccessfully exhausted, check whether gl, is a root node. If it is, STOP, else do: * let gi < gk* go to step 2. The algorithm identifies a set of candidate goals that get proposed iteratively in an attempt to break the deadlock. The weaker goals, the ones with lower contributions to the higher level goals, are proposed first.The PERSUADER uses the strategyof presenting candidate goals in terms of increasing strength so as to take into account that the cumulative effect of presentinga seriesof reformulationswill also be instrumentalin changing the parties' perceptions. Focus of attention is provided by the rejected goal. The search starts with siblings of the rejected goal and expands to searching for higher-level goals to introduce. 6. Goal Substitution Besides case-based reasoning, goal substitution is performed in the PERSUADER by searching the goal graph of the recalcitrant party. The input to the process is the goal or set of goals that have been the cause of negotiation deadlock. Each rejected goal is used as a starting point in searching the goal graph of the rejecting party. The algorithm for goal substitution is basically the same as the one for goal introduction. Let gi be the goal causing the current deadlock. The algorithm proceeds as follows: 1. Match gi in the goal graph. 2. Traverse the contribution link to gi's parent goal gk,. 3. Traverse the links downwards to the set T of children of g,, (siblings of gi). 4. Out of the set identified in step 3, delete goals that are currently present. Out of the remaining goals select the ones whose contribution to gk.is greater than or equal to the contribution of gi. This set constitutes the candidates to be substituted for gi.
" "Match"means matchingthe framerepresentations of the objectinstances.If there is more than one
objectionable issue, use the algorithm for each one. The objectionable goal can always be found in the goal graph since it is one of the issues under negotiation. 12 gi could have more than one parent goal. If this happens, the algorithm is applied for each of the parents.

1264

KATIA P. SYCARA

The control strategy the PERSUADER uses for selecting goals to be substituted for gi out of the elements of T is to rank order the set of substitution candidates in terms of contribution to gk,,delete goals already present and out of the remaining ones, substitute the goal that has the smallest contribution first.If this substitution does not help eliminate the deadlock, each of the remaining goals in the candidate set is tried in ascending order of contribution to g,. If all substitution candidates have been exhausted without success, let gi - g, and repeat the algorithm. The algorithm terminates when g, is a node with no outgoing edges. 7. Goal Abandonment There are two strategies implemented in the PERSUADER that result in goal abandonment by a party: (a) promise the party the fulfillment of a more important goal if he abandons the current goal, (b) point out that insistence on the current goal threatens a more important goal of the party. For example, the promise of higher wages can make a union abandon its goal of modification of management's rights clauses (strategy (a)).13 The threat of losing employment can make a union abandon its goal of wage increases (strategy (b)). Which are the goals targeted for abandonment? What is the knowledge and algorithms that enable a computer system to automatically generate threats and promises that result in abandonment of goals that hinder movement towards agreement? The goals that are targetedfor abandonment are the issues on which a party has refusedto make concessions. The knowledge that the PERSUADER uses is the models of the negotiating parties represented in the goal graphs. Figure 4-3 has shown a partial belief graph of a company. Figure 7-1 depicts a partial belief graph of a union. The processingdepends on which party must be induced to abandon goals. To convince a union to abandon one of its goals using thr-eats,the strategy is to search for company actions that threaten one of the union's important goals. To convince a company, the strategy is to discover whether the company's refusal will result in a violation of an

UNION-PRESTIGE

(+)

MGT-RIGHTS (-) EMPLOYEE-SATISFACTION

UNION-SECURITY (+)

(+)

EMPLOYMENT (+) NON-ECONOMIC (+) ECONOMIC (+)

SENIORITY (+)

SUBCONTRACT (-)

WAGES (+)

FRINGES (+)

JOB-SECURITY (+)
FIGURE 7-1.

Partial Belief Structure of a Union.

13 The Coors Beer company has used this tack with their nonunion employees to avoid weakening management rights;

PROBLEM RESTRUCTURING IN NEGOTIATION

1265

important company goal. To convince a union to abandon a goal using promises the algorithm is similar to the one for threats except that instead of finding a more important persuadee goal that is violated, the algorithm searches for a more important persuadee goal that is desired. A goal is violated by an action if the action opposes its sign. For example, if the company lays off employees, a reduction in employment, EMPLOYMENT(-), occurs, violating the union's goal EMPLOYMENT(-+). Since the company controls the hirings, firingsand concessions, both of these strategiesrequirea goal directed a set of posearch of the company s goal graph. Company goal graph search pr-oduices tentially appropriate threats. Search of the union's goal graph is used for selectionl of appropriate threats out of the generated set. In general, if one party controls the contingencies, then its own goal graph needs to be searched to generate threatening arguments. Let us assume that party-2 controls the contingencies and party-I does not. The algorithm to produce a threat to convince party-I to abandon issue X ( *) (where * denotes a (+) or a (-) in the parties' goal graphs) is as follows: 14 1. Find out which of party-2's goals are violated by party-I's refusal. This is done by following the contribution links startingwith X(NOT*) in party-2 goal graph, i.e., tracing the consequences for party-2 of the negation of its goal. The effects of negating X are propagated by changing the signs of X's ancestor goals along the path. 2. Find out what compensating actions party-2 might carry out to offset the effects of negating its goal X. This is done by considering the subgoals ZI, * * *Z, of each goal Y found in step 1. To qualify as a threatening argument, a potential compensating action Zi has to satisfy three conditions: (a) it must be controllable by party-2, (b) it must violate a goal of party-I and (c) the importance for party-I of this violated goal must be greater than the importance of the demand under discussion. If the third condition is not satisfied by Zi, its subgoals are checked to see whether they satisfy conditions (a) to (c); otherwise, the subgraphconsisting of Zi and all of its contributing subgoals is pruned, and the sibling goals of Zi are considered in the same way. If some Z,, proves suitable, a potential argument has been found. Whether or not an argument has been generated, steps I and 2 are repeated starting from Y and generating a set of potential arguments. The algorithm at worst visits all nodes in the graph in a Depth-First fashion. Visited nodes are marked, so they are not revisited. The complexity of Depth First Search is O(j El), linear in the number IEl of edges (Even 1979). Generating an argument to convince party-2 about issue X is similar: the X(NOT *) path is followed in party-2's goal graph. The deleterious results that X(NOT*) has on one of its higher level goals are pointed out to party-2. The algorithm to produce a promise to convince a persuadee to abandon a goal is similar to the algorithm for producing threats. The difference lies in step 2 (a). To qualify as a promise, a compensating action Zi, instead of violating a goal of the persuadee, has to be desired by the persuadee, i.e., have a (+-) sign in the persuadee's belief structure. Moreover, in step 2, there is an additional condition (d) that a potential compensating action Zi must satisfy: (d) the cost to the persuader of effecting the promised goal must be less than the cost of the abandoned goal. When the argument-generatingprocess described above produces more than one potential threat, the best order of presentation must be chosen. One strategy is to try the "weakest" argument first, presenting "strong" arguments only if the weaker ones fail. This requiresmeans of ranking arguments according to their "convincing" power (Sycara 1990a). The ranking follows the order of importance of the goals that the arguments threaten. In particular, the importance of the goals of a company (union) depends on
14 For clarity of presentation, the algorithms in termsof two parties.If morethanone partyhas we present

rejected a proposal, conditions (b) and (c) of argument acceptability become: (b) it must violate goals of all rejecting parties, and (c) the importance for eachl rejecting party of the violated goal must be greater than the importance of the demand under discussion.

1266

KATIA P. SYCARA

the financial situation of the company, the state of the industry, the labor supply and the general economic climate. For example, the goal of reducing labor cost is more important for a company in an industry with high labor cost; if there is abundant labor supply in an area, the goal of employment is stronger for a union in that area. In this case, a threat of layoffs has the greatest convincing power for the union. Consider a negotiation between a company and its union, where the PERSUADER has suggested a compromise. The company has agreed saying that the increase in fringe benefits (fringes) is the highest it can afford. The union wants a higher increase. Deadlock occurs. The PERSUADER restructuresthe problem by attempting to make the union abandon the goal of higher increase in fringes. This is done by producing the threat "If the company is forced to grant higher increases in fringes, then it will decrease employment". To generate the above argument, the PERSUADER finds out which company goals are violated by the union's refusal. Then, it finds out what compensating actions the company might use in retaliation. To do this, the PERSUADER matches the fringes goal in the company's belief graph (see Figure 4-3). It propagates the increase in fringes that the union wants to enforce to the ancestors of the fringes goal (ECONOMICCONCESSIONS, LABOR-COST, PRODUCTION-COST, PROFITS). Children of these ancestors (e.g., WAGES, EMPLOYMENT) might indicate subgoals that the company can fulfill to counteract the increase. Such a counteracting action (e.g., EMPLOYMENT(-)) that violates a union goal (EMPLOYMENT(+)) (see Figure 7-1) that is more important than the increase in fringes constitutes a threat that is aimed at making the union abandon the goal of further increases in fringes. 8. Changing Reservation Prices If other restructuring methods have failed, the PERSUADER attempts to change the reservation price of the objecting party. Changing reservation prices is difficult since it involves changing not simply the payoff of a proposed settlement but changing the threshold of acceptability itself. Besides using case-based reasoning for this task, changing reservation prices is performed through the generation of threats to break off negotiations if the party does not accept the current proposal. Breakoffarguments assert that the cost of breaking off negotiations is greater than the loss in payoff associated with the current (unacceptable to the rejecting party) proposed settlement. The effectiveness of these threats reflects actual labor negotiations (e.g., Herman and Kuhn 1981 ) where it is observed that as the strike deadline approaches, and as the parties calculate the costs of a strike against the possible worth of a settlement, their willingnessto compromise increases, or equivalently, the payoff required for agreement decreases. To calculate the loss in payoff of the current unacceptable proposal, the PERSUADER 1 In other uses a heuristic acceptability threshold of 70% payoff for each of the parties. words, it is assumed that a party will accept a proposed settlement if it gives payoff (equal and/or) higher than 70%.The loss in payoff is the difference between 70%and the payoff of the current unacceptable proposal. The loss in payoff also represents the amount by which the reservation price of the rejecting party needs to be changed so that the party will agree to the proposed settlement. In the current implementation, it is assumed that loss of payoff is proportional to the party's breakoff costs. In other words, the greaterthe negotiation breakoff cost, the greater the payoff loss the party can tolerate (the less the needed payoff of an acceptable proposal). In the labor negotiations domain, breakoff costs are strike and lockout costs. To calculate breakoff costs for the union and the company, the factors taken into consideration include: the inventory situation of the company, the economic conditions in the industry, outstanding orders, the condition of
15 This

was suggested by our two domain experts.

PROBLEM RESTRUCTURING IN NEGOTIATION

1267

the union's treasury, unemployment in the job classifications of the bargaining unit and the strike history of the union. For a target reservation price change, the PERSUADER calculates breakoff cost and checks to see whether it is greater than the targeted change. If it is, this calculation forms the basis for a negotiation breakoff argument. If it is not, the PERSUADER seeks to see what assumptions would increase the breakoff cost and the breakoff arguments include these assumptions. For a detailed discussion of breakoff arguments, see Sycara ( 1987). In certain situations, such as labor negotiations, the breakoffcost is a function of time. In such domains, a persuader has to convince the rejecting party that negotiations will not be resumed for as long a time as is needed for the negotiation breakoff cost to yield the targeted payoff drop. In the PERSUADER this length of time is looked up from the curve of strike cost vs time for the rejectingparty. This curve is approximatedby accessing experiences with similar parties. The shape of a company's curve depends on the kind of product the company produces. For example, if the grape pickers go on strike at the time the grapes are ripe, there is an initial period of a few days where the strike cost rises linearly with time (assuming the strike started just before the optimal picking time). After the grapes are overripe, the strike cost drops to zero. For the union, the curve of strike cost vs time depends on the shape the union's strike fund is in, the level of unemployment for the skills of the union members, and the level of unemployment benefits. 9. Concluding Remarks Negotiation is a dynamic process where the interactions among the parties change the space of feasible solutions. A crucial aspect of negotiations, especially in distributive negotiations where the parties do not in general share goals, is problem restructuringto generate a new problem representation. The problem restructuring process facilitates expansion of the space of feasible settlements and the generation of possibly commonly held goals, thus unblocking negotiations. Restructuring is a highly creative process that depends on the experience and insights of the negotiators. It is also nondeterministic since it must take into consideration the constantly evolving negotiation context. Automating the restructuringprocess is a challenging task that explores the limits of current Al technology. We have presented a framework and mechanisms to effect computerized problem restructuringin negotiations. The presented algorithms have been implemented in the PERSUADER, a computer program that acts as an automated labor mediator in hypothetical negotiations. The restructuring processes the PERSUADER uses are: ( 1) case-based reasoning, (2) situation assessment, (3) goal graph search and (4) persuasive argumentation (in particular generating threats and promises). Although the PERSUADER uses knowledge from the domain of labor relations, the restructuringprocess and algorithms are domain independent.'6
16 This

researchwas funded in part by the Army Research Office under contract No. DAAG 29-85-K-00230.

References
The Bureau of National Affairs, Washington, DC, ELKOURI, F. AND E. ELKOURI, How Arbitration W-Vork-s, 1972. Computer Science Press, Rockville, MD, 1979. EVEN, S., GraphAlgorithmns, GOELTNER, C., "The Computer as a Third Party," Technical Report, Sloan School of Management, M.I.T., Cambridge, MA, Working paper, (1987). Sci., 35, 7 (1989), 788-806. GUPTA, S., "Modeling Integrative Multiple-Issue Bargaining," Mcanagement HERMAN, E. AND A. KUHN, Collective Bargaininlgan-dLabor Relationis, Prentice-Hall, Inc., Englewood Cliffs, NJ, 1981.
JARKE, M., M. T. JELASSI AND M. F. SHAKUN, "MEDIATOR: Towards a Negotiation Support System,"

European J. Oper. Res., (1987).


KEENEY, R. L. AND H. RAIFFA, Decisionis with AuldtipleObjectives, John Wiley and Sons, New York, 1976.

1268

KATIA P. SYCARA

KERSTEN, G. E., L. BADCOCK, M. IGLEWSKI ANDG. R. MALLORY, "Structuringand Simulating Negotiation: An Approach and an Example," TheorYanldDecision, 28, 3 (1990). J. L., Retrievaland OQganizational KOLODNER, Strategies in Conceptuial Memory:A Comiipuiter Lawrence Mfoclel, Erlbaum Associates, Hillsdale, NJ, 1984. , R. L. SIMPSON AND K. SYCARA-CYRANSKI, "A Process Model of Case-Based Reasoning in Problem Solving," In Proc. Nintil Joint Inter;iat. Conf. Artificial Intelligence (IJCAI-85), Los Angeles, CA, 1985, 284-290. M. F., Evoliutioniary and Grolup Decision Slupport SHAKUN, UnclerComnplexitv Systems Design: Policy AMaking Systems, Holden-Day, Oakland, CA, 1988. SYCARA, K., "Resolving Adversarial Conflicts: An Approach IntegratingCase-Based and Analytic Methods," PhD thesis, School of Information and Computer Science Georgia Institute of Technology, 1987. ,"Utility Theory in Conflict Resolution," Ann. Oper. Res., 12 (1988), 65-84. ,"Persuasive Argumentation in Negotiation," Thleoryand Decisions, 28 (1990a), 203-242. ,"Negotiation Planning: An Al Approach," EulropeanJ. Oper. Res., 46 (1990b), 2 16-234.

Вам также может понравиться