In this chapter we have shown several contributions made toward automated negotiation, and also its classifications, applications, research problems, and some active
research areas. Our research in this thesis is only concerned with a branch of automated negotiation: bargaining.
Figure 2-5. Context of this thesis in automated negotiation and some of its related work
Another purpose of this chapter is to show the context of this thesis among existing research in terms of three major dimensions (protocol design, application areas, problem solving approaches including theoretical and empirical analysis). As we will see later in the next chapters, it is concerned with bargaining protocol design supported by
APPROACH:
Theoretical Analysis - Game theoretic approach - Computational approach
- Argumentation-based negotiation Application Areas
- E-commerce Negotiation Property
- Resource allocation - Competitive
- Task allocation - Cooperative
- Education/training
[Sandholm and Vulkan, 1999] (strategic delay) [Faratin, 2000] (open MAS)
[Larson and Sandholm, 2002] (strategic delay) [Sim and Wang, 2004] (strategic delay)
[Karunatillake and Jennings, 2005]
(avoiding argumentation)
[Faratin et al., 1998] (decision function) [Faratin, 2000] (decision function, simulation) [Fatima et al., 2001] (simulation)
[Sim and Wang, 2004] (simulation)
both theoretical and empirical analysis. Figure 2-5 shows the topics and approach used in this thesis and its relation to other work in automated negotiation.
The topics and approaches concerned are highlighted in Figure 2-5. The closest work to this thesis is Faratin’s work on building an agent’s decision function in an uncertain environment (assuming the agent does not know information about its opponents). Later, his approach is elaborated in [Fatima et al., 2001] and [Sim and Wang, 2004]. Our agent’s decision function in the non-monotonic bargaining case is also derived from Faratin’s work, with further extension to the strategic delay and strategic ignorance cases. The framework will be described in Chapter 3.
Sandholm and his colleagues [1999, 2002] use a game theoretic approach to analyze bargaining with a deadline. The results show the benefit of strategic delay, which is similar to our work. However, they did not verify the benefit empirically due to a restrictive domain of their study. Sim and Wang also propose a strategic delay under a specific condition, which is tested empirically. However, they did not provide a general agent decision framework using strategic delay. This thesis provides a general agent decision framework on deciding a delay and empirical results in terms of cost and benefit of using strategic delay in bargaining.
Karunatillake and Jennings [2005] briefly analyze the cost of argumentation by means of empirical study. They suggest a withdrawal by an agent from a negotiation when the argumentation in that negotiation is costly. Our work is different in the sense that we do not suggest a withdrawal, but rather to employ strategic ignorance on costly topics. In addition, we also propose a decision function for invoking strategic ignorance among agents and using simulator to show the benefit of it.
Finally, since this thesis discusses three different modifications to bargaining protocols, most related work for each modified protocol will be discussed separately in Chapter 4 when the modifications are introduced. Some other related work, which is relevant to the general framework of our agent’s decision model, will be provided in Chapter 3.
CHAPTER 3
MODELLING AGENTS AND THEIR BARGAINING STRATEGIES
The strategy used by an agent in bargaining depends on the information and computational capacity of the agent. Without considering any information, the simplest strategy of a buyer is to submit consecutive offers based on a predefined sequence, e.g.
<$100, $110, $125, …, $200>. Using this sequence, the buyer starts with an initial offer of $100, which then is revised to $110, $125, and so on until it is either accepted by the seller or reaches the ceiling $200. Since the sequence is set by the user and the agent only executes it regardless of the negotiation situation, we classify it as a static strategy. This method was commonly used in early systems, such as Kasbah [Chavez and Maes, 1996].
The advantage of this strategy lies in its simplicity and transparency for the user.
However, it restricts the agents from taking advantage of relevant information including their opponent’s behavior. Consequently, it tends to work well only when the bargaining information is known by the user, or when uncertain factors are not important in determining the success of the negotiation. For example, if the buyer knows that the seller will accept a $200 offer and it does not have a time deadline, then the aforementioned sequence can certainly be used by the buyer, even though it does not guarantee a maximum surplus.
In more sophisticated strategies, agents may have a higher autonomy to decide their offers, as long as they are not exceeding the limit set by the user. This is more likely to be used when the bargaining involves many uncertain factors. In these strategies, the offers may change according to the information sensed by the agents; thus, they are referred to as dynamic strategies. In this chapter, several dynamic strategies are discussed that may be implemented in bargaining under uncertainty. The structure of the discussion of these strategies presented in this chapter is shown in Figure 3-1.
Figure 3-1. Topics covered in Chapter 3
First, we briefly discuss the characteristics of the bargaining problem and solution under uncertainty. Then we look at agents’ behaviors, starting with their evaluation criteria in making a decision (section 3.2.1), followed by a description of agents according to their negotiation strategy (section 3.2.2). Then, we propose a model of agent belief and its revision mechanism, which is an important part of myopic agents, i.e.
agents that only consider a few periods ahead in their decision making. More attention is Dynamic Strategies
under Uncertainty (3.2)
Evaluation Criteria (3.2.1)
Myopic Agents (3.2.2)
- Maximize expected gain - Myopic-0
- Myopic-1 Bargaining Problem
and Solution (3.1)
Agent’s Belief (3.3) Model and revision
- EvalF-I agents - EvalF-II agents
Discussion (3.4) Summary (3.5)
given to myopic agents and their belief construct because they are used as the basis of the theoretical analysis, which makes this analysis different from traditional game-theoretic analysis that assumes a perfect forecast. After describing the model, we discuss the implication of the model in section 3.4. Finally, the summary of Chapter 3 is provided in section 3.5. All concepts described in this chapter will be used as the basis for the analysis of the modified bargaining protocols proposed in the next chapter.