3.2 Future challenges
3.2.2 Alternative decision making algorithms
Existing approaches for determining the software agents’ strategies ignore inherent features of the World Wide Web, as an open medium for automated negotiation, where software agents for automated negotiation can easily be programmed by human users. This openness – for new software agents of user with various preferences and for many different negotiation problems – of media for automated negotiations has to be considered when designing decision making algorithms for software agents.
Evolutionary computing needs many repetitions of negotiations over one negotiation problem in the best possible environment – consisting of opponent software agents as realistic as possible – to achieve competitive strategies for automated negotiation by means of co-evolution (Beam and Segev, 1997). However, when automated negotiations are conducted in the World Wide Web, the number of negotiations with one specific opponent might be quite low and opponents and negotiation problems may change from one negotiation to the other. Moreover, novel software agents might enter the market and new negotiation problems could appear. Therefore it is un- likely that a software agent has many trials against one opponent for one negotiation problem, which causes a major disadvantage for evolutionary computing in developing the agents’ deci- sion making algorithms for automated negotiations (Bichler, 2000). Especially strategies based on sequential threshold rules will have problems to cope with different opponents if they are optimized against only some opponent strategies and preferences (Tu et al., 2000).
While learning software agents are designed to cope with opponents with various preferences in learning them, the possibility of novel software agents also causes problems for them. The model of the opponent, learning agents hold in mind to learn the parameters of this model from the course of negotiation with the opponent, will probably be inadequate to model the diversity of possible existing as well as novel opponents. Furthermore the learning capacities are severely limited by the small number of parameters learning agents can effectively process in automated negotiations.
Also time-based strategies for automated negotiation have their drawbacks, though the incor- poration of time in strategies for a basically dynamic processes like negotiation is in general appreciable. In traditional negotiations time is likely to have a major impact on the process and outcome of negotiations, as stated by Cross (1965, p.72): ’As any economist knows, time has
a cost, both in money and in utility terms; it is our position that it is precisely this cost which motivates the bargaining process. If it did not matter when people agreed, it would not matter whether or not they agreed at all.’ However, it is questionable whether this statement is valid in
automated negotiation that reach agreement – or otherwise terminate without an agreement – in some seconds only. The fast and low cost proceeding of automated negotiation is one of the possi-
3.2. Future challenges 67
ble sources of comparative advantage over alternative dispute resolution mechanisms, besides the assumed better outcome and novel transactions it enables. Furthermore it is arguable whether designers of software agents should base their decision making algorithms for software agents in automated negotiation on behavioral heuristics – as mentioned above, time-based strategies build on the observation of human negotiation behavior in experiments – given the aim of such agents to improve the outcome of negotiations over that reached by humans through automating the negotiation process. Interestingly simulation studies revealed that time-based and imitating strategies reached more agreements and agreements of higher utility in negotiations with later deadlines (Faratin et al., 1998). This raises the question of why to impose a deadline at all when the decision to terminate negotiations could be delegated to the software agents as well. Finally, imitating and trade-off strategies do not face the problems mentioned so far, but are for their own insufficient to determine a complete decision making algorithm for a software agent. Imitating strategies can decide the extent of a negotiation step (concession, demand, or insistence) in terms of the difference in utility between two subsequent offers of the software agent. They can easily derive this difference from this difference of their utility between the two previous offers of the opponent. However, this reciprocation of the concession magnitude provides no guidelines for the final configuration of an offer, which is a problem when there exist some offers between which the user is indifferent. In contrast to imitating strategies trade-off strategies have problems to determine when and how much to concede or demand, as they are designed to determine the package configuration for a given utility level. Changes in the utility level of offers, however, will at some point in the negotiation be necessary to avoid getting stuck when no more offers of the same utility level are available.
Given these deficiencies of existing approaches for determining the the decision making algo- rithms of software agents in automated negotiation, continuous concession strategies proposed in negotiation literature seem to be a viable alternative. Continuous concession strategies follow rule-based deterministic concession algorithms that neither model their opponent nor try to learn something about their preferences or strategy, but are therefore (re)usable for various negotiation problems with different and novel opponents. Furthermore, in not making offers dependent on time, they also respect the low transaction costs associated with automated negotiation in the World Wide Web. The continuous concession strategies proposed in Chapter 4 are not novel ones, but were partly proposed in negotiation literature and not yet implemented in simulation studies on automated negotiation. These strategies are similar to the ’dummy’ strategies used for comparison purposes in some of the studies reviewed above, however more sophisticated. Though these mechanisms seem very simplistic, this plainness also has its advantages. First, for accep- tance of software agents it must be clear to human users how they decide and act to establish trust and thereby enhance their application in practice (Nwana et al., 1998; Maes et al., 1999). Second, not the complexity of software agents but its performance counts. In a problem closely related to negotiation – the iterated prisoner’s dilemma – the tit-for-tat strategy, though being the simplest – measured in number of code lines – repeatedly outperformed opponent strategies in tournaments (Axelrod, 1980a,b).
Obviously it is important to investigate the performance of the agents’ strategies when negotiating against each other, for different negotiation problems and in different interaction protocols, to determine which one yields the best outcomes. However, in the decision whether to use automated negotiation instead of traditional negotiation between humans the benchmark for the evaluation
of software agents is the outcome of human negotiation. To automate negotiations in substituting their human users, or replace other currently used transaction mechanisms, software agents have to achieve better outcomes (Blecherman, 1999).17 Therefore it is not only necessary to use
preferences as close as possible to human preferences as input for the software agents – as advocated in the previous section –, but it is also necessary to compare the results of automated negotiations between software agents to the results of traditional negotiations between humans (or the currently used transaction mechanisms) – for the same preferences and therefore the same negotiation problems – to evaluate the performance of software compared to human agents. Even if one strategy outperforms all the others it will not be applied in automated negotiations if it fails to outperform human negotiators, unless saved transaction cost compensate for this deficiency.