Evolutionary Ideals

6.7.1 Reciprocation, Evolution, Trust

In a society of agents, interactions between agents will proceed apace. Indeed, it is likely that two agents will meet more than once, if the society is small enough, or if the two agents are working in the same area, geographically or otherwise. In such interactions, it is an evolutionary strength to be a reciprocal cooperator (Harcourt, 1991). In other words, an agent,a, will do better if he reciprocates help with another, b, than if he did not. Reciprocation is a common form of behaviour in the animal world (Harcourt, 1991; Trivers, 1985). The great apes help each other in fights, for example, in the hope (or knowledge) that their status is increased in the troupe, and that help will come their way from those they help (Trivers, 1985). Vampire bats feed those who have not eaten well one night, in the knowledge that they in turn will be fed on a bad night (Harcourt, 1991). It follows that, if the bats who fed others were not in their turn fed, they would die out, and so would their altruistic behaviour. They are fed, however, because this apparent altruism enables the population as a whole to survive, and ensures the survival of individual bats.

Reciprocal altruism is a form of trust in the animal world (Harcourt, 1991). In more intelligent animals capable of reason, reciprocation is more common if some form of trust as we understand it is present. If trust was not present, the initial risks would not be taken to be altruistic (Rempel et al., 1985; Boon & Holmes, 1991), reciprocation would not occur, and the collective society would be worse off. To illustrate this, consider the following example.

If two agents,aand bwere reciprocating trusters, and another two,cand d, were not disposed to cooperation at all, and if they were members of a team which hunted by night, the following situation may hold. On any one night, there will be the chance that any of the agents may not get enough food. In terms of payoffs, this situation is given the value 0. If an agent gets enough food for itself, its payoff is 3, and if it shares

this food, the payoff to each agent sharing the food is 2. We consider two consecutive nights. On the first, a does well, b does not, and a shares its catch. Payoff to each = 2. On the same night, cdoes well,d does not, but crefuses to share its food, thus payoff to c is 3, to d, 0. On the next night, b does well whilst a does not, and so, since bis a reciprocator, it shares its food witha. The cumulative payoff for each is 4. Likewise, ddoes well, c does not, and so the cumulative payoff for each of these two, since drefuses to share its food, is 3. So despite the short term gainscandd make by not reciprocating or trusting, the long term result is a fitter a and b, and this fitness can translate into leadership of the troupe, for example. Reciprocation can pay off in such a circumstance.

The above example is somewhat contrived, but serves to illustrate the point that long term reciprocation is good for individuals and the group as a whole (Harcourt, 1991; Trivers, 1985). Of course, on a good night, all agents get a payoff of 3. There will be at some time, however, a bad night for some of the agents, and it is then that the differences arise. A run of bad nights fordwill result in its death from starvation, and a famine will result in the deaths of non-reciprocators, whilst those who reciprocate and are altruistic may survive.

This general principle can be made use of in the formalism, particularly with regard to the alteration of trust values. See the previous chapters, also chapter 8, for further discussion. In summary, reciprocation is more likely in a trusting relationship (Boon & Holmes, 1991) and reciprocation is good in evolutionary terms (Harcourt, 1991).

6.8

Summary

This chapter presented and discussed several ‘principles’ which trust, or at least the formalism for trust presented in the previous chapter, appears to follow. In some senses, the principles themselves are of limited importance. What is important is that the formalism provides the means to discuss such principles precisely and clearly, and with much less ambiguity than has previously been the case.

That trust can be discussed with little ambiguity is one of the goals of the thesis. Another is that the formalism can be implemented and embedded in an artificial reasoning agent. The following chapter presents discussions and implementations of the formalism, with the aim of proving the utility of both the formalism and trustper sefor such agents.

Practical Work

7.1

Introduction

There have been several investigations into the concept of trust in the past. They ranged from experiments involving questionnaires (Rotter, 1967; Rempel & Holmes, 1986) to more abstract visions based on philosophy, sociology, social psychology, and other social sciences (Barber, 1983). None of them provided a usable formalism for trust, and very few attempted a workable explanation of the concepts involved. As far as is known, none have thus far provided an implementation of their ideas.

The present work goes beyond other explorations of trust by providing two major contributions:

1. A workable formalisation;

2. A simple implementation to demonstrate the concept.

The first of these was discussed in earlier chapters. The remainder of this chapter concentrates on the second, but is limited in its realisation. It is, however, an indica- tion of what is possible, and an important confirmation that the formalism introduced here is workable and can be embedded in artificial agents.

The experiments described below were in fact carried out several times, but the amount of data generated by each was quite large. It was decided to concentrate on and report typical examples of each experiment with the aim of showing the general style of results for each.