In this chapter we have introduced the essential concepts that underpin the work presented in this thesis. Cooperation and coordination are key processes for increasing the aggregate welfare of agent societies, and in this thesis we focus on two major mechanisms for encouraging such behaviour: trust and reputation (Chapter 3) and norms and conventions (Chapters 4, 5, 6, 7).
Nowak’s rules for cooperation (discussed in Section 2.1.2) provide a useful analytical starting point for describing the mechanisms behind the promotion of cooperative behaviour, and underpin the processes for both major mecha- nisms investigated in this thesis. A number of research questions remain open for both trust and reputation, and norms and conventions. Firstly, in Section 2.2 we discuss the need for supporting cooperation through indirect reciprocity. This raises questions regarding the accuracy of reputation assessments, partic- ularly in the context of open MAS domains, and we investigate this in Chapter 3. Secondly, we discuss in Section 2.3.2 limitations with current theories of con- vention. Effectively applying mechanisms for norm and convention emergence requires resolution of these limitations, since it is highly likely that a single convention will not be an attainable goal in large open MAS. We discuss this in Chapter 4, and subsequently identify further directions for research which underpin the rest of the thesis.
This chapter should be seen as a broad introduction. Where relevant, we include further detailed background discussion of these concepts throughout this thesis. For example, Chapter 3 discusses trust and reputation in more detail and Chapter 4 provides a detailed analysis of research into convention emergence. We have provided an overview of network concepts and background literature in Appendix A for those unfamiliar with the field.
Trust, Reputation and Gossiping
As discussed in Chapter 1, trust and reputation are fundamental mechanisms for protecting individuals from selfish or malicious behaviour in open MAS. In such systems, we can expect complex network structures and extremes of information availability (i.e. there may exist systems in which agents have insufficient infor- mation and systems in which agents have incomplete information, since they cannot feasibly observe all occurring interactions). In this chapter, we use a simple model of reputation to investigate the effects of incomplete information and underlying network structure on levels of cooperation in a population. We show that insufficient or incomplete information can undermine the efficacy of reputation and allow selfishness to dominate. We apply a simple gossiping algo- rithm to supplement observation of agent behaviour and show significant drops in levels of selfishness in the population.
3.1
Introduction
Many typical approaches to increasing levels of cooperative behaviour in highly decentralised open MAS domains have involved biasing interactions towards cooperative individuals. Such mechanisms serve two purposes: (i) protecting agents from individuals likely to engage in selfish behaviour and (ii) increasing the aggregate welfare of the population. The structure of many MAS domains implicitly creates incentives for selfish behaviour, such as free-riding in BitTor- rent and other P2P networks (Ruberry & Seuken, 2012), or energy conservation
in wireless sensor networks (Galstyanet al., 2004).
Trust and reputation mechanisms, which incorporate observations and indi- vidual experience to aid decision making, introduce this bias into agent partner selection through direct (trust) and indirect (reputation) reciprocity (Nowak & Sigmund, 2005). An agent who has cooperated in the past is more likely to receive reciprocal cooperation from others. In domains in which the identity of interaction partners is known, trust and reputation can facilitate significant increases in aggregate welfare, but their efficacy is directly related to the quality and quantity of information available about individuals in the population (Som-
merfeldet al., 2008). Trust, which is based on direct observations of behaviour,
can only be effective once historical interaction data are available. Reputation, which relies on observation or propagation of third-party agent behaviour, may
be undermined byincomplete information, in which agents make decisions based
on unrepresentative sets of observations. Direct and indirect reciprocity involve feedback effects: a cooperative action can cause many subsequent cooperative actions, and vice versa. Consequently, decisions made on incomplete informa- tion may be incorrect, in the sense that given full information the agent would have acted otherwise, and these mistakes will be amplified by the feedback of reciprocity.
Network topology also plays a significant role in the dynamics of trust and reputation mechanisms. By definition, agents are constrained to interact only
with their direct neighbour set, and interaction behaviour can only be observed by those directly connected. Networks can support isolated communities of cooperators (Nowak & Sigmund, 2005) and the role of network structure in
facilitating information propagation is well studied (e.g. Glinton et al.(2010),
Newman (2003), Wang (2003)).
Trust and reputation mechanisms are highly suited to open MAS domains, and present a useful setting for investigating the impacts of incomplete informa- tion and network structure on levels of emergent cooperative behaviour. In this chapter, we empirically analyse the conditions under which mechanism efficacy
is reduced and demonstrate a possible mechanism, namelygossiping, to mitigate
the effects of incomplete information and exploit the ease of information trans- mission in typical network structures. Specifically, we show that incomplete information can result in inaccurate reputation assessments that subsequently reduce cooperation, and that the underlying network structure significantly in- fluences emergent behaviour, both positively and negatively depending on the configuration. We supplement trust and reputation with gossiping, which can be used as a substitute for direct observation of interactions and has a low space and time complexity, using one of four aggregation rules. We show that gossip- ing can reduce selfishness in the population by up to 25%, and is particularly effective on real-world networks.