• No results found

Distributed Artificial Intelligence

DAI is a research field which is growing in applicability and relevance, both to Ar- tificial Intelligence and to computing as a whole. In short, it involves concepts of distribution, intelligence, society, independence, graceful degradation, and localised decision making. It is, then, an ideal tool for the study of trust. The following sec- tion summarises the strengths of DAI as a tool for the study of an implementation of trusting behaviour in artificial agents, taking into account the criticisms of other approaches discussed above.

3.5.1 The use of DAI as a Research Tool

With independent, rational intelligent agents, we are presented with the ideal tool for researching implementations of trust. We summarise the reasons here:

Agents are assumed to be (pseudo-) intelligent. Since in societies of agents, intelligent behaviour must include considerations of the beliefs and attitudes of others who may affect the agent who is considering, trust presents an agent with an extra capability in that direction. In addition, without seeking to anthro- pomorphise, trust as an affector of behaviour in artificial agents may present interesting viewpoints on human agents.

Agents are assumed to be (pseudo-) rational. Whether or not trust is rational depends on many things. In Deutsch’s investigations, the story of the Princess

20Whilst emergent behaviour can arise in DAI, or any other complex system. Also, DAI, as a

superset of AL, can make use of reactive aspects of agents (Chapman & Agre, 1987; Spector & Hendler, 1990; Downs & Reichgelt, 1990)

21It is granted that, in animals, trusting behaviour, such as in the form of delayed reciprocation, can

and does take place (Harcourt, 1991). Indeed, in this work, we take such an open view of trust. The fact remains, however, that the human race makes most efficient use of the phenomenon, to the extent that it has become a part of our everyday existence, and we deliberate about it, both consciously and unconsciously.

and the Tiger presents an interesting view of this, with some of the forms of trust (e.g., trust as despair, trust as faith) not necessarily being rational, and others (trust as confidence) being rational in many cases. For rational agents, a goal would be the implementation of rational trust. Trust as confidence supplies this viewpoint. However, instances of irrationality in this area of trust can be studied more closely in supposedly rational agents.

Agents are generally assumed to be cooperative (von Martial, 1992). In fact, agents are generally assumed to be trustworthy also, and clearly neither assump- tion can be justified in the ‘wider world’ that exists outside the research labs. Nevertheless, there are situations within which cooperation is necessary (Con- nah & Wavish, 1990; Marsh, 1992), and others where, although cooperation may be necessary, there is a choice about who to cooperate with (Marsh, 1992). In both of these types of situation, the behaviour of the formalism can be tested and refined.

Agents are distributed. Trust is a societal concept (Luhmann, 1979; Baier, 1986), and distributed agents, with non random interactions between them (ei- ther through choice or direction) are a society in themselves (Bond & Gasser, 1988; Gasser, 1991). Thus trust is not out of place and can be observed in interactions within such artificial societies.

Agents are generally independent. Trust is a means of coping with the freedom of others (Gambetta, 1990a; Luhmann, 1979). Within a society of independent agents, its behaviour can be readily investigated and observed, and anomalies detected.

There is a wide range of potential and actual applications for DAI, from Air Traffic Control (Cammarataet al., 1983) to Open Informations Systems (Hewitt, 1991). Trust can potentially be implemented and observed in realisations of many such spheres, and its behaviour and influence detected and refined over time.

3.5.2 Amalgamation

The implementations to be discussed in chapter 7 are not strictly DAI. In fact, they are very closely allied to Game Theory in that they use the Prisoners’ Dilemma situation, but are an amalgamation with many of the concepts of DAI in that the agents are independent, social, and geographically distributed, with some control of movement. Thus, the agents are limited in terms of intelligence, although they are knowledgeable of the aspects of the Prisoners’ Dilemma, particularly in terms of payoffs and utility, and so they are more complex than the usual entities of Artificial Life. Future work (see chapter 8) envisages more detailed intelligent trusting agents.

3.6

Summary

Trust is an issue which is close to all members of society, however it is used (Luhmann, 1979). It is thus an extremely complex area, and one which is not properly understood

or researched (Luhmann, 1990). This chapter has presented a summary of many of the prevailing views of trust, from fields such as sociology, social psychology, and economics, to name but some. It is inevitable, with trust being so all-pervading, that such a review cannot cover the entire sphere in detail, and will obscure some aspects of that field. What was attempted then was to present the view of trust that several more prominent researchers in the field held, comparing and contrasting these views. The chapter then discussed aspects of trust which, although considered of importance, were not easily classified in these terms. The result is a comprehensive, although not complete, review of the work done concerning trust, and the thoughts of several prominent minds. A complete review is both outside the scope of the present work and unnecessary in the attempt to provide an understanding of the concept.

The present work is concerned with formalising the concept in a way that en- ables its inclusion in artificial agents. Several methods for examining such an idea were critically discussed, with the conclusion that the field of DAI presented unique opportunities in the form of pseudo-rational, intelligent, sometimes cooperative, in- dependent agents. Such agents, it is argued, will also benefit from the inclusion of a concept of trust in their decision-making repertoires.

To that end, the following chapters present and work with the formalism that has been developed in this work. Chapter 4 presents the formalism, based on an amalgamation of much of what has been presented here, and refined much over the past three years. Chapter 5 discusses the formalism, showing examples of it at work. Chapter 6 uses the formalism in a descriptive way, discussing some principles for trust. Chapter 7 culminates with discussions of implementations of the formalism in simple artificial agents. Such implementations not only follow the correct experimental path for work of this kind, but also justify the claim that the formalism can be implemented and produce agents whose trusting behaviour is as can be expected.

An Example Heuristic

Formalism

“Justice is represented by a square number”

Pythagoras. “. . .values are not the least vague when you’re dealing with them in terms of actual experience.”

Pirsig, 1991, page 63.

4.1

Discussion

From the previous chapters, it is clear that a unified theory for trust is lacking. In order to develop such a theory, and associated principles, it is necessary to provide a precise means of discussion about the phenomenon. Deutsch (1973, developed further by Golembiewski and McConkie (1975)) can be seen as starting to develop such a means of discussion in the various aspects of trust that he brought together into one solid consideration of the subject. The present work provides a formalisation 1 of trust

with which to continue the discussion in a precise, unambiguous manner. As was suggested in chapter 2, the introduction of a new formalism has potential problems, not least that the formalism will be seen by many as too restrictive, particularly with something as rooted in our subconscious and unconscious thoughts as trust. Acknowledging those problems, we argue that the formalism as it stands can help in discussions about them, leading to better versions of itself.

This chapter introduces the formalism that has been developed. The early devel- opment of the formalism was largely based on discussions of trust in the literature (see chapter 3). However, the later work built upon the initial formalism using experimen- tal results and other observations, including discussions with others whilst applying it to areas other than cooperation in DAI (e.g., as in Thimbleby et al., 1993). What follows, then, is a formalism that is considerably more applicable than when it started

1There is perhaps confusion between the use of the words ‘formalisation’ and ‘formalism’. For the

remainder of the thesis, we discuss a ‘formalisation of trust’ in terms of a ‘formalism for trust’. The two are considered to be interchangable here.

out (and as presented in Marsh (1992), Marsh and Thimbleby (1992)) without losing any of the intuitively appealing aspects of trust that it first described, such as the three forms of trust (see below). The formalism consequently has more applicabil- ity to the consideration of the phenomenon of trust in social situations than it first achieved.

The main consideration in the introduction of such a formalism is which approach to take. In this work, the aim was to provide a formalism which was as simple to understand as possible, whilst preserving its expressive power (the use of Occam’s Razor). One reason for this was to allow an implementation based very closely on the formalism to be developed. This consideration led to the decision to use simple probabilistic methods for describing trust. This argument is put in chapter 2. A differing approach to the concept of trust — that of harmony, or Wa — is discussed in chapter 8.

4.1.1 Overview of the Chapter

The formalism, as might be expected of something describing trust, is large in the sense that extensions are possible, have been described, and continue to be made. This chapter presents the formalism ‘as is,’ in the hope that it will spur others to extend it and correct it further, and the knowledge that such work is indeed possible. To allow a greater understanding of the formalism and associated formulæ, we present it in a stepwise fashion. Thus the next section will present the initial aspects of the consideration, describing how we represent trust at all in a formalism. Following that, we present the major aspect of the formalism — its application to agents deliberating in a potentially cooperative situation. Acknowledging the restrictions in the formulæ, we proceed to extend them, attaching other considerations, finally achieving a solid formulation which considers temporal constraints, differing situations and the simi- larities between them, and the past behaviour of other agents and the environment as a whole (although such a consideration is, in this chapter, implicit). Chapter 5 extends the formalism further, showing its use for describing considerations of trust, and chapter 6 presents some principles which trust in general adheres to.