Bootstrapping trust and stereotypes with tags

In real-world environments, cooperation often emerges amongst agents who are observably similar. Estimating the expected behaviour of another agent is a challenging problem, particularly for new agents who have little or no experience of others. In this paper, we show how observable features can be used to find similar, and hence cooperative, partners. Our contribution extends trust and stereotype approaches, to include comparisons and learning of observable features, called tags. In environments where no reciprocity exists (or where there have been insufficient interactions for reciprocity to take effect) tags have been used to encourage cooperation. The only information available to an agent early in its life is knowledge of its own tags and behaviour. We assume that agents who are observably similar will be behaviourally similar too. Agents use reinforcement learning to take advantage of as much available information as possible, until sufficient experience has been gathered for more established trust and stereotype models to be built. Our results show that using tags improves agents’ rewards in the early stages of their lifetime when used prior to established stereotype and trust algorithms. We demonstrate that tags are successful in supporting cooperation, even when agent behaviour is independent of the partner, because the approach correctly identifies similar agents. Good agents are able to select partners who will act as they do, while bad agents avoid those who are observably similar.