HISTORIA DE LAS REVISIONES
3. DESCRIPCIÓN GENERAL DE LA ACTIVIDAD
There are two major implementations that have been carried through for this work. Both are concerned with the Iterated Prisoner’s Dilemma (PD) (Axelrod, 1984; Behr, 1981; Dolbear & Lave, 1967). There are reasons for this choice.
1. The PD is a well-known, well-understood ‘game,’ in the social sciences and in computing. Consequently, any results or insights gained in the experiments carried out here will have a wide audience and applicability.
2. Being constrained, the PD provides us with the ideal tool for experimenting with agents since, from the agent’s point of view, the only choices available to them or the agent with whom they are interacting are cooperation or defection. Such constraints allow the agent to reason in a limited fashion, whilst still providing us with some measure of the result of such reasoning.
3. Because the PD is so limited, it follows that simple trusting agents can easily estimate costs, benefits, and utility for a situation for themselves and also for those with whom they interact. In the future in more complex interactions, such considerations will be agent-subjective and estimated by the agents themselves. 4. Payoffs in the PD present an ideal measure of how well particular agents, and indeed the society as a whole, are performing. This measure can be used to as- certain the relative successes of particular strategies as a whole or in interactions with other strategies (or themselves), and could be used in a form of ‘survival of the fittest’ to show the relative evolutionary stability of particular strategies (Lomborg, 1992).
In short, the PD provides us with a seemingly ideal tool for performing research and fine tuning trust strategies. There are drawbacks to its use, however:
1. The PD is by its very nature confrontational. This must be the case in its classic form, otherwise no dilemma exists, because “noncooperative short-run maximising behaviour is inconsistent with long-run (cooperative) behaviour” (Dolbear & Lave, 1967). Agents must naturally enter a PD interaction in a confrontational state of mind — hence the dilemma. For experiments in trust as we see it, this is an inherent drawback since we see trust as an avenue for
achievingcooperation where it may not seem possible. The PD makes this more difficult.
2. Because of the nature of the PD, in order to ‘do well,’ i.e., score more points, or increase fitness, one has to defect before one’s opponent (Behr, 1981). This is unfortunate, but since it is a phenomenon that is well known, can be taken into account when performing experiments.
3. Objections to the PD have been raised in other areas, notably by Argyle (1991), who lists the following differences between the PD and everyday life:
(a) Play is simultaneous, with ignorance of the move of the other player, and there is a risk inherent in the other’s failing to cooperate. Argyle suggests that in real life, if a person decides not to do something, their payoff is 0. Thus, if A decides not to play tennis or join in house building, his payoff is again 0. Whilst this is not strictly the case, we leave this argument for the present.
(b) The game is too abstract. In this instance, this is of little importance. (c) There is usually no communication. In situations where communication
takes place, cooperation is generally increased (Argyle, 1991; Deutsch, 1962). Since the PD generally disallows communication, this may hin- der its results. In the case of the artificial agents used here, this is not
too much of a problem, since we do not expect communication to produce bargaining behaviour at present.
(d) The players are usually strangers, and are invisible to one another. This is not always the case in the following implementation.
(e) Social norms are absent. Cooperative or defective norms are absent since the PD, being abstract, is a game without such rules. In the case of the fol- lowing implementation, again, this has little significance since social norms are learned from childhood (Argyle, 1991), and our agents exist ‘as is,’ with no remembered ‘childhood,’ although such a consideration could be made in the future.
Despite these objections, the PD provides a constrained, simple, yet realistic mirror of the kind of interactions that we would expect trusting agents (or any artificial agent) to be involved in. Such interactions are likely to be characterised by many of the same things that the PD is characterised by, such as:
1. Each interaction has specific ends (to increase payoff, to gain information, to enlist help) and agents are aware of these ends (or can at least make an ‘educated’ guess about them.
2. The respective payoffs are readily known, or can be estimated.
3. The other agent may or may not be known. If known, then memory can help in the decision making process.
4. The result of an interaction is generally clear cut (cooperate or not, give infor- mation or withhold it, give or withhold help), although for artificial agents this is less true in many situations — the giving of information, for example, can be partial, or somewhat clouded.