3. TECTÓNICA
3.2. DESCRIPCIÓN DE LA ESTRUCTURA
3.2.2. Las fases de deformación
In this section we summarize the aims and objectives of this thesis and the contributions to the state of the art that were made to achieve them. Our general aim is to develop techniques that help to reduce the uncertainty in repeated multi-agent negotiations in open distributed systems. To this end, we set out to achieve the following particular objectives:
• Develop a comprehensive trust model for multi-agent systems that can evaluate the reliability or honesty of agents.
• Use trust in bargaining encounters and mechanism design in order to reduce the uncertainty agents have about their negotiation opponents’ reliability and hon- esty. This involves using the trust measure developed by our trust model in an agent’s reasoning mechanism (i.e. in its bargaining strategy) and developing the protocol for an interaction mechanism that caters for the uncertainty regarding the reliability of agents.
• Develop a comprehensive model of persuasive negotiation that comprises: (i) a pro- tocol that incorporates the use of arguments and determines what commitments hold whenever agents make offers or issue arguments and (ii) a reasoning mech- anism that can generate offers and arguments and can evaluate and respond to these during a negotiation encounter. This requires developing both an argument generation and evaluation component, as well as thestrategies for PN.
• Demonstrate the benefit of using arguments in automated negotiation and show that they enable agents to reach agreements more efficiently (i.e. better and faster) than using normal negotiation protocols that only allow an exchange of offers.
• Implement PN in a realistic context in order to demonstrate its applicability and effectiveness in managing inter-agent dependencies.
To achieve these objectives, a number of contributions were made to the state of the art:
• In (Rahwan et al., 2003b), we provided the first survey of the state of the art in the area of ABN and identified the main trends and challenges that pervade the field. This survey set the landscape within which we develop our model of PN and appears as chapter 2 in this thesis.
• In (Ramchurn et al., 2004b), we provided a critical analysis of the trust issues that arise in MAS. In particular, we showed how various models developed in MAS form a coherent approach towards resolving uncertainties about the reliability and honesty of agents. Thus, we also identify the current challenges in the field which we aim to meet in our model. This review appears in chapter 3 in this thesis.
• Based on our preliminary work in (Ramchurn et al., 2004d) and the requirements presented in chapter 3, we describe a novel trust model (called CREDIT) that enables an agent to measure its opponents’ trustworthiness (honesty or reliabil- ity) over multiple encounters (Ramchurn et al., 2004c). The model is shown to be effective and efficient at preventing exploitation by opponents by allowing the agent to adjust its negotiation stance in repeated bargaining encounters accord- ing to its trust in its opponents (hence reducing uncertainty). Moreover, using CREDIT’s trust measure, an agent is also able to select its interaction partners more effectively. CREDIT is presented in chapter 5.
• Given our work on CREDIT, we then introduced the use of trust modelling to the area of mechanism design by developing the notion of Trust-Based Mechanism Design (TBMD) to reduce uncertainty about the reliability of agents (Dash et al., 2003). In so doing, we created the first efficient, individually rational, and incen- tive compatible mechanism that takes into account the trust agents have in each other. This is, in effect, the first efficient reputation mechanism that incentivises agents to reveal their impressions of others truthfully. Specifically, our Trust-Based Mechanism combines these measures into an overall trust measure (using a trust model such as CREDIT) to select those agents that are best at doing certain tasks. This work is presented in chapter 6.
• While CREDIT and TBMD are concerned with uncertainties about reliability and honesty, in (Ramchurn et al., 2003a) we provided a preliminary model of Persuasive Negotiation whereby agents can use threats and rewards to elicit better agreements by reducing uncertainties about preferences and action sets. This model describes the general concepts that are used to develop our new PN mechanism that can allows agents to reach better agreements faster than standard bargaining mech- anisms. In particular, we develop a new protocol and reasoning mechanism for agents to use to give or ask for rewards in repeated encounters. We also show, em- pirically, that agents are able to engage in more efficient and effective agreements using this protocol and reasoning mechanism than only bargaining with offers. We also develop a novel strategy for PN and show that it enables agents to achieve even better agreements than current negotiation strategies. The complete model is given in chapter 7.
• Given our model of PN, we then apply it, together with CREDIT, in a pervasive computing environment (Ramchurn et al., 2004a). In so doing, we are able to show, for the first time, how a PN and trust model can be used in practice to allow agents to resolve their conflicts effectively. In particular, in this work we show how PN can be used by agents to negotiate about the usefulness of interruptions in a meeting room scenario. In so doing, negotiating agents provide an effective way to reduce the intrusions caused by interruptions and help their human owners to focus on the main task undertaken during the meeting. This work appears as chapter 8.
Drawing all these together, the application of the various models we develop to cater for uncertainties in negotiation is graphically expressed in figure 1.3. As can be seen, CREDIT and TBMD overlap in that they deal with the uncertainty regarding the re- liability and honesty of agents. In CREDIT, we show how to develop and use trust in bargaining encounters, while in TBMD we show how to use the core concepts of CREDIT in mechanism design. Given this, CREDIT also overlaps with persuasive negotiation as they both apply to bargaining encounters and both try to reduce uncertainty about the action set of agents. They do this by either adjusting the negotiation stance of an agent (i.e. the selection of values for issues in this context) or by using arguments to con- strain the action set. In addition to this, PN overlaps with TBMD since PN also aims to explore preferences more efficiently than standard bargaining techniques through the use of arguments while TBMD aims to elicit these preferences through the protocol it enforces upon the agents together with the trust model it uses.
Incentive Compatibility Efficiency Individual Rationality Protocols Strategy Negotiation Object Commitments Issues Trade-offs Games Game
theory Heuristics ParticipationRules
Preferences
Environment
(action set) Trustworthiness
= The Cloud of Uncertainty
Mechanism Design
Bargaining
Reliability
= Region where TBMD applies
= Region where PN applies
= Region where CREDIT applies TBMD
Negotiation
PN
CREDIT
Figure 1.3: Applying CREDIT, TBMD, and PN to reduce the uncertainty underlying negotiation.