Image scoring in ad-hoc networks: An investigation on realistic settings Henry Franks, Nathan Griffiths, and Arshad Jhumka Department of Computer Science, University of Warwick, CV4 7AL, UK Abstract. Encouraging cooperation in distributed Multi-Agent Systems (MAS) remains an open problem. Emergent application domains such as Mobile Ad-hoc Networks (MANETs) are characterised by constraints including sparse connectivity and a lack of direct interaction history. Image scoring, a simple model of reputation proposed by Nowak and Sig- mund, exhibits low space and time complexity and promotes cooperation through indirect reciprocity, in which an agent can expect cooperation in the future without repeat interactions with the same partners. The low overheads of image scoring make it a promising technique for ad- hoc networking domains. However, the original investigation of Nowak and Sigmund is limited in that it (i) used a simple idealised setting, (ii) did not consider the effects of incomplete information on the mech- anism’s efficacy, and (iii) did not consider the impact of the network topology connecting agents. We address these limitations by investigat- ing more realistic values for the number of interactions agents engage in, and show that incomplete information can cause significant errors in decision making. As the proportion of incorrect decisions rises, the efficacy of image scoring falls and selfishness becomes more dominant. We evaluate image scoring on three different connection topologies: (i) completely connected, which closely approximates Nowak and Sigmund’s original setup, (ii) random, with each pair of nodes connected with a con- stant probability, and (iii) scale-free, which is known to model a number of real world environments including MANETs. 1 Introduction The emergence of cooperation remains an open problem in the agent commu- nity. Although significant progress has been made in a number of domains, few proposed techniques are fully applicable given the challenges of emergent do- mains such as MANETs. These domains are characterised by unique constraints such as low computational capacity, high agent turnover, lack of centralised au- thority, sparse connectivity and lack of single ownership, and therefore require low-overhead, distributed solutions. Investigating such mechanisms for promot- ing cooperation has become a major theme of research in open MAS. Nowak and Sigmund’s [16] image scoring mechanism biases partner selection towards cooperative individuals while requiring low space and time overheads. Agents maintain a subjectively perceived image score of other agents, based on