State 3: The agent is known and trusted
8.6 Other Aspects
The formalism presented in this thesis is a starting point for work concerned with trust, particularly in Multi-Agent Systems. It is argued that, an agent who can reason with and about trust will be more capable of making reliable, informed decisions than one who simply ignores the phenomenon, or assumes it is present all the time, or, worse, who assumes it is not at all present. Since it is a starting point, however, it leaves much to be done. The previous sections have suggested further work in such areas as sociology, DAI, HCI and CSCW, in addition to mentioning certain specific augmentations of the theory of trust we have built up. In this section, we present
further enhancements to or changes to the formalism as it stands, discussing various problems which are evident, and suggesting ways to address them.
8.6.1 Why Use Values?
Chapter 2 discussed the use of values in this formalism. It was argued that, whilst explicit values placed on trust pose certain problems (e.g., whose subjective, or agent- centered, value we are observing, and whether or not two agents’ values that are the same mean the same things in terms of trust) the use of values does carry certain benefits. Not least of these benefits is that we are able to provide a formalism which is linear and is simple enough to understand intuitively whilst preserving the power of both the formalism and the concept of trust. In addition, we are able to experiment with different value bases, to allow us to determine which base is most representative. We are not alone in the use of values — Gambetta (1990a) has suggested that trust takes a value in the interval [0,1] (see chapter 3). That said, as far as we can ascertain we are the first to provide a formalism for use with trust, and the first to suggest an implementation of that formalism.
The use of a different value base provides some interesting questions. If trust were suggested to take a value in the interval suggested by Gambetta, no real sensitivity would be lost, since we use real numbers. So, anything in the range [0,0.5) maps to our value of [−1,0), and any value in the range [0.5,1) would take a value in the range [0,1) in our formalism. The formalism as it stands would probably have to change somewhat. One of its problems is inherent in the use of fractional numbers multiplied by other fractional numbers. Inevitably, the result is smaller than the multipliers. Whilst this works well when we consider that we have negative values to represent degrees of distrust, thus anything positive, however small, represents some degree of trust. It does not work at all when we are restricted to positive values for all degrees of trust and distrust. There are two major problems with the more restricted range of values:
1. It is less clear that we are talking about trust or distrust at any one time. 2. The formalism needs to be considerably more sophisticated to take trust and
distrust into account.
The first point is to a large extent academic, since the artificial agents using trust would not notice the difference. The second point is more problematic. One solution is to convert the more restricted range to its mapped value in the larger range, perform the calculation, and map the result back. This is clearly impractical. A modification to the formalism is required. This is beyond the scope of the current work, and would need to be devised over a long period of time and experimentation. It is unfortunate that the formalism cannot be generalised across different value ranges. This problem can be addressed in the future, should the present range of possible values for degrees of trust become inadequate. This is unlikely, since any such range is infinite. In practice, however, a more large-grained range may be more useful to artificial agents. Since it remains to be seen if a complete formalism of the workings of human trust can be found, this may be largely an academic debate.
A final aspect of the use of values concerns their sign. Clearly, multiplying negative values results in positive values. This is an unexpected bonus with regard to the formalism, particularly with the formula for situational trust:
Tx(y, α) = Ux(α)×Ix(α)×Tdx(y)
With a negative utility estimated, it follows that an agent would not wish to carry through this situation. If a negatively trusted agent was present, the final trust value would be positive (assuming importance is positive). Thus the distrusted agent is
trusted positively not to carry out whatever task is needed in this situation. This is an interesting anomaly, but on consideration seems useful. 2
This interpretation of trust is clearly a Machievellian approach to trust. Naturally there are others. These other approaches, for example masochistic, are dependent on the algebra chosen to represent the formalism. Different algebras may well result in different approaches. See section 9.6.
Deceit
Despite all of the checks and balances supplied by the law, safety nets, and experience with other agents, there will still be the possibility that agents can and will be deceitful. Here we present a brief discussion which raises some interesting questions about trust and deceit.
In practice, if the workings of the formalism are common knowledge, then any agent can make use of this to manipulate the truster into a position where they can be taken advantage of by behaving in a manner which actually increasestrust in the short run, then reversing this behaviour in order to gain the maximum benefit for themselves at the expense of the truster. Interestingly enough, this can be seen in trust in humans, hence the success of confidence tricksters.
How can we reinforce trusting agents against such events? The intuitive answer is that this is not possible — if trust and its workings are common knowledge, then deceitful behaviour is possible. The workings of the formalism aside, there may be a method in the algorithm we use for the alteration of trust following specific behaviour. It is generally accepted (see earlier) that trust increases following positive behaviour, and decreases following negative behaviour. The amount of the increase or decrease is what is of issue here. As suggested by Rempel et al (1985, 1986), a sudden defection from a trusted friend can result in a drastic reduction of trust, to the extent that a lot of work is necessary to build that trust up again. This suggests that confidence tricksters can only get away with their tricks once with any one person. This does not help society as a whole, however, until the information about that trickster is disseminated. In humans, emotions such as shame or embarrassment may hinder such dissemination. We suggest that in artificial agents, this would not be the case, and that, coupled with the use of the enforcer, sanctions could be taken immediately the trick came to light. Thus, at the cost of one agent, society is protected against such strategies. Note that, as trust, such an answer can only exist in societies of agents, where communication is available between agents.
To take deceit into account, the formalism may not need to be extended, but more thought has to be put into the alteration of the trust values used by agents. One of the limitations of what has been presented in this work is that of the alteration of trust. Indeed, this is a large topic deserving much further work. The testbed developed for this work provides the ideal ground for experimentation in this area.