Opponent Selection and Concurrency Control

Concurrency control means selecting the number of negotiations to engage in con-currently and opponent selection refers to choosing opponents for each of these negotiations. All the papers discussed in this section assume that the agent can negotiate with all possible providers. Although this may be possible in some situa-tions, we do not think that this is generally the case and we assumed that resources to be used in negotiation are limited (requirement R4). On the other hand, our sellers are going to be heterogeneous (requirementR3), as probably are the sellers in most markets. For these reasons, it is therefore necessary to choose the number of opponents and the actual opponents to negotiate with. Since these issues have not been addressed in the concurrent bilateral negotiation literature, we have to look elsewhere.

Unfortunately we could not find any literature on concurrency control, nor could we specify any other problem, which would have been discussed in the literature and would be similar enough to be useful for our purposes. However, as we ex-plained, concurrency control is important. Also it matters because of a risk of getting into too many contracts. In some environments (cheap/free decommit-ment), it might not be a problem, but where the decommitting is very expensive,

a useful buyer agent should be more careful. For this reason, we will develop meth-ods for concurrency control in our work. Using decision theory seems a reasonable starting point for that work, because it allows us to consider different aspects of the situation in a unified way.

We could not locate any literature on opponent selection either, but we were able to find another problem that is close enough and has plenty of literature on it.

This other problem is service selection. In the service selection problem, an agent has a certain task to be performed and has to choose services that satisfy its needs optimally. However, there are two differences between opponent and service selection:

• Dependencies: Instead of finding an optimal service to fulfil a certain task, the service selection task is usually performed over several interconnected tasks and the goal is to find the services that are interoperable and produce the best possible result. Since we assume that our negotiations can have this type of connections, this might actually be quite useful, although at this level we are only interested in selecting the best service for the one task at hand. We will return to the interconnectedness later. However, since it is the dependencies between services that make the service selection an interesting problem, the actual selection of one service usually plays a very minor part in these papers.

• No negotiation: The work of service selection usually assumes that there is no negotiation on the price or quality, but that these have been set by the providers on a take-it-or-leave-it basis. Therefore the price is usually known in advance exactly and quality is an estimate (because it can vary).

However, for us it is other way around: each provider has a fixed quality (but in some case may decommit and therefore not perform at all) but the price is a result of negotiation and therefore, it is uncertain.

With these differences in mind, we now discuss the literature. Service selection is usually based on multiple criteria, so the literature uses many standard techniques for such situations, such as multi-attribute utility theory (Keeney and Raiffa 1976) and decision theory (Raiffa 1987). Also some other approaches have been used.

We will now discuss how these could be applied to our problem.

The basic idea of multi-attribute utility theory is that the overall value of a service consists of the values of its performance on the relevant characteristics, for example

its quality and reliability. Usually, simple additivity is assumed, so that the value of a service is equal to:

v(s) =

k

X

i=1

wivi(s),

where vi(s) is a normalised value of the characteristic i for service s, k the number of characteristics, and wi is the relative importance of that characteristic for the decision-maker. Normalisation means that given any possible value of characteris-tic i the relevant value function Vi has a value in the interval [0, 1]. The better the attribute, the higher the value; however this relation does not need to be linear but can take any form. Another common assumption is thatPk

i=1wi = 1. Now, it is clear that the critical task in multi-attribute utility theory is to find the appro-priate weights for the overall value function. There are many standard techniques for deriving these, probably the best-known being the analytic hierarchy process (Saaty 1980). The service selection papers that use some version of multi-attribute utility theory include Seo et al. (2004) and Zeng et al. (2003).

However, in situations where the utility is quite clearly not a sum of characteristics, but something else, multi-attribute utility theory does not work too well. Here, we defined the consumer’s utility as: U (s) = V (qs) − p, where V (qs) is the value of service s and p the price (one of the characteristics). Now, if we also consider the provider’s reliability ws, the probability that it will provide the agreed service at the agreed time, we get:

Uconsumer(s) = ws[v(qs) − pc] + (1 − ws)f ee,

where f ee is the decommitment fee the provider has to pay if he fails to deliver the service. Now, the different characteristics are no longer additive, so the basic form of multiattribute utility theory is not applicable.

In contrast, in the decision theory approach the decision-maker will simply use his utility function to calculate his expected utility for using each possible service and then he just chooses the option that provides the highest expected utility (Collins et al. 2001).⁵³ Here, the expected utility would be calculated using the equation for utility and replacing ws, qs and pc with the values from the service in question.

53The expected utility is of course very strictly confined to the situation at hand (as discussed in section2.2.1.1) and since the situation of Collins et al. is significantly different (they have a reverse auction and multiple interconnected services), it is not useful to go through their model in detail. Instead, we will discuss some problems of using this approach in our environment.

However, although each provider’s quality of service qs and reliability ws are as-sumed to be known (requirement R6) and also the decommitment fee f ee and the consumer’s value function v(..) can be assumed to be known by the consumer agent, the problem is the price pc, which is not usually known at the time of op-ponent selection by anyone, since it will be determined by a negotiation after the opponents have been selected. The provider-specific price distributions are also likely to be very hard to learn even over considerable time, since they depend on the negotiation tactics used by the parties and their negotiation parameters (such as deadlines and reservation prices), which are likely to vary between negotiation encounters. However, we can make this choice properly when we use negotiation tactics that make the same offer again and again. Then we know exactly what the price is going to be if we are successful in the negotiation and a contract is formed. Of course we need to consider and be able to estimate the probability of being successful in a negotiation, z, so we get the following:

Uconsumer(s) = z(ws[v(qs) − pc] + (1 − ws)f ee).

In order to get a useful estimate to z, we would need to know what negotiation tactic the opponent is using or at least we need to know the possibilities and the probability distribution over them. If we have very little or no information about the opponent tactics, the meaning of this extension will be very small. If z is roughly the same for the all the opponents, it will make a very small impact on the selection. Despite these problems, the decision theoretic approach seems applicable to our problem and so we will adopt it (see section10.1.2.1for details).⁵⁴ We will of course use also other (much simpler) opponent selection techniques, but we use the decision theoretic approach as a pinnacle of opponent selection, the most sophisticated technique.

Using a decision theoretic approach in both opponent selection and concurrency control means that we will need a significant amount of information about the market and the sellers. Our information requirements (requirement R6) have made it possible for us to consider and choose decision theoretic approaches to some issues, although such information may not always be readily available. We,

54There are also other approaches to service selection, such as the preference ranking organi-sation method for enrichment evaluations (PROMETHEE) (Brans and Vincke 1985), which was used for service selection bySeo et al.(2005), and constraint satisfaction/optimisation (used for service selection in Lin et al. (2005)). They all suffer from the same problem: the best goal function depends on pc, which may be difficult to estimate in advance and they are not directly applicable to our problem. The PROMETHEE is quite complicated and there are very few constraints in our setting.

however, think that it is good approach to see first what works when there is good information and then make the setting more challenging (and realistic) by taking some of that information away and trying to cope with less information.

We also think that although our information needs are significant, they can, with some additional work, often be relaxed with a limited loss of performance. We also use less sophisticated methods in some cases to see how they perform. Of course this work should be continued in future work to settings with less and less information.⁵⁵